# source:doc/papers/concurrency/Paper.tex@f1f8e55

aaron-thesisarm-ehcleanup-dtorsdeferred_resndemanglerenumforall-pointer-decayjacob/cs343-translationjenkins-sandboxnew-astnew-ast-unique-exprnew-envno_listpersistent-indexerresolv-newwith_gc
Last change on this file since f1f8e55 was 5ff188f, checked in by Peter A. Buhr <pabuhr@…>, 4 years ago

further changes to document Makefiles

• Property mode set to 100644
File size: 146.7 KB
Line
2% red highlighting ®...® (registered trademark symbol) emacs: C-q M-.
3% blue highlighting ß...ß (sharp s symbol) emacs: C-q M-_
4% green highlighting ¢...¢ (cent symbol) emacs: C-q M-"
5% LaTex escape §...§ (section symbol) emacs: C-q M-'
6% keyword escape ¶...¶ (pilcrow symbol) emacs: C-q M-^
7% math escape $...$ (dollar symbol)
8
9\documentclass[10pt]{article}
10
11%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
12
13% Latex packages used in the document.
14\usepackage[T1]{fontenc}                                        % allow Latin1 (extended ASCII) characters
15\usepackage{textcomp}
16\usepackage[latin1]{inputenc}
17\usepackage{fullpage,times,comment}
18\usepackage{epic,eepic}
19\usepackage{upquote}                                            % switch curled '" to straight
20\usepackage{calc}
21\usepackage{xspace}
22\usepackage[labelformat=simple]{subfig}
23\renewcommand{\thesubfigure}{(\alph{subfigure})}
24\usepackage{graphicx}
25\usepackage{tabularx}
26\usepackage{multicol}
27\usepackage{varioref}
28\usepackage{listings}                                           % format program code
29\usepackage[flushmargin]{footmisc}                              % support label/reference in footnote
30\usepackage{latexsym}                                           % \Box glyph
31\usepackage{mathptmx}                                           % better math font with "times"
32\usepackage[usenames]{color}
33\usepackage[pagewise]{lineno}
34\renewcommand{\linenumberfont}{\scriptsize\sffamily}
35\usepackage{fancyhdr}
36\usepackage{float}
37\usepackage{siunitx}
38\sisetup{ binary-units=true }
39\input{style}                                                   % bespoke macros used in the document
40\usepackage{url}
42\usepackage{breakurl}
43\urlstyle{rm}
44
45\setlength{\topmargin}{-0.45in}                         % move running title into header
47
48%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
49
50% Names used in the document.
51
52\newcommand{\Version}{1.0.0}
53\newcommand{\CS}{C\raisebox{-0.9ex}{\large$^\sharp$}\xspace}
54
55\newcommand{\Textbf}[2][red]{{\color{#1}{\textbf{#2}}}}
56\newcommand{\Emph}[2][red]{{\color{#1}\textbf{\emph{#2}}}}
57\newcommand{\R}[1]{\Textbf{#1}}
58\newcommand{\B}[1]{{\Textbf[blue]{#1}}}
59\newcommand{\G}[1]{{\Textbf[OliveGreen]{#1}}}
60\newcommand{\uC}{$\mu$\CC}
61\newcommand{\cit}{\textsuperscript{[Citation Needed]}\xspace}
62\newcommand{\TODO}{{\Textbf{TODO}}}
63
64
65\newsavebox{\LstBox}
66
67%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
68
69\setcounter{secnumdepth}{2}                           % number subsubsections
71% \linenumbers                                          % comment out to turn off line numbering
72
73\title{Concurrency in \CFA}
74\author{Thierry Delisle and Peter A. Buhr, Waterloo, Ontario, Canada}
75
76
77\begin{document}
78\maketitle
79
80\begin{abstract}
81\CFA is a modern, \emph{non-object-oriented} extension of the C programming language.
82This paper serves as a definition and an implementation for the concurrency and parallelism \CFA offers. These features are created from scratch due to the lack of concurrency in ISO C. Lightweight threads are introduced into the language. In addition, monitors are introduced as a high-level tool for control-flow based synchronization and mutual-exclusion. The main contributions of this paper are two-fold: it extends the existing semantics of monitors introduce by~\cite{Hoare74} to handle monitors in groups and also details the engineering effort needed to introduce these features as core language features. Indeed, these features are added with respect to expectations of C programmers, and integrate with the \CFA type-system and other language features.
83\end{abstract}
84
85%----------------------------------------------------------------------
86% MAIN BODY
87%----------------------------------------------------------------------
88
89% ======================================================================
90\section{Introduction}
91% ======================================================================
92
93This paper provides a minimal concurrency \textbf{api} that is simple, efficient and can be reused to build higher-level features. The simplest possible concurrency system is a thread and a lock but this low-level approach is hard to master. An easier approach for users is to support higher-level constructs as the basis of concurrency. Indeed, for highly productive concurrent programming, high-level approaches are much more popular~\cite{HPP:Study}. Examples are task based, message passing and implicit threading. The high-level approach and its minimal \textbf{api} are tested in a dialect of C, called \CFA. Furthermore, the proposed \textbf{api} doubles as an early definition of the \CFA language and library. This paper also provides an implementation of the concurrency library for \CFA as well as all the required language features added to the source-to-source translator.
94
95There are actually two problems that need to be solved in the design of concurrency for a programming language: which concurrency and which parallelism tools are available to the programmer. While these two concepts are often combined, they are in fact distinct, requiring different tools~\cite{Buhr05a}. Concurrency tools need to handle mutual exclusion and synchronization, while parallelism tools are about performance, cost and resource utilization.
96
97In the context of this paper, a \textbf{thread} is a fundamental unit of execution that runs a sequence of code, generally on a program stack. Having multiple simultaneous threads gives rise to concurrency and generally requires some kind of locking mechanism to ensure proper execution. Correspondingly, \textbf{concurrency} is defined as the concepts and challenges that occur when multiple independent (sharing memory, timing dependencies, etc.) concurrent threads are introduced. Accordingly, \textbf{locking} (and by extension locks) are defined as a mechanism that prevents the progress of certain threads in order to avoid problems due to concurrency. Finally, in this paper \textbf{parallelism} is distinct from concurrency and is defined as running multiple threads simultaneously. More precisely, parallelism implies \emph{actual} simultaneous execution as opposed to concurrency which only requires \emph{apparent} simultaneous execution. As such, parallelism is only observable in the differences in performance or, more generally, differences in timing.
98
99% ======================================================================
100% ======================================================================
101\section{\CFA Overview}
102% ======================================================================
103% ======================================================================
104
105The following is a quick introduction to the \CFA language, specifically tailored to the features needed to support concurrency.
106
107\CFA is an extension of ISO-C and therefore supports all of the same paradigms as C. It is a non-object-oriented system-language, meaning most of the major abstractions have either no runtime overhead or can be opted out easily. Like C, the basics of \CFA revolve around structures and routines, which are thin abstractions over machine code. The vast majority of the code produced by the \CFA translator respects memory layouts and calling conventions laid out by C. Interestingly, while \CFA is not an object-oriented language, lacking the concept of a receiver (e.g., {\tt this}), it does have some notion of objects\footnote{C defines the term objects as : region of data storage in the execution environment, the contents of which can represent
108values''~\cite[3.15]{C11}}, most importantly construction and destruction of objects. Most of the following code examples can be found on the \CFA website~\cite{www-cfa}.
109
110% ======================================================================
111\subsection{References}
112
113Like \CC, \CFA introduces rebind-able references providing multiple dereferencing as an alternative to pointers. In regards to concurrency, the semantic difference between pointers and references are not particularly relevant, but since this document uses mostly references, here is a quick overview of the semantics:
114\begin{cfacode}
115int x, *p1 = &x, **p2 = &p1, ***p3 = &p2,
116        &r1 = x,    &&r2 = r1,   &&&r3 = r2;
117***p3 = 3;                                                      //change x
118r3    = 3;                                                      //change x, ***r3
119**p3  = ...;                                            //change p1
120*p3   = ...;                                            //change p2
121int y, z, & ar[3] = {x, y, z};          //initialize array of references
122typeof( ar[1]) p;                                       //is int, referenced object type
123typeof(&ar[1]) q;                                       //is int &, reference type
124sizeof( ar[1]) == sizeof(int);          //is true, referenced object size
125sizeof(&ar[1]) == sizeof(int *);        //is true, reference size
126\end{cfacode}
127The important take away from this code example is that a reference offers a handle to an object, much like a pointer, but which is automatically dereferenced for convenience.
128
129% ======================================================================
131
132Another important feature of \CFA is function overloading as in Java and \CC, where routines with the same name are selected based on the number and type of the arguments. As well, \CFA uses the return type as part of the selection criteria, as in Ada~\cite{Ada}. For routines with multiple parameters and returns, the selection is complex.
133\begin{cfacode}
134//selection based on type and number of parameters
135void f(void);                   //(1)
136void f(char);                   //(2)
137void f(int, double);    //(3)
138f();                                    //select (1)
139f('a');                                 //select (2)
140f(3, 5.2);                              //select (3)
141
142//selection based on  type and number of returns
143char   f(int);                  //(1)
144double f(int);                  //(2)
145char   c = f(3);                //select (1)
146double d = f(4);                //select (2)
147\end{cfacode}
148This feature is particularly important for concurrency since the runtime system relies on creating different types to represent concurrency objects. Therefore, overloading is necessary to prevent the need for long prefixes and other naming conventions that prevent name clashes. As seen in section \ref{basics}, routine \code{main} is an example that benefits from overloading.
149
150% ======================================================================
151\subsection{Operators}
152Overloading also extends to operators. The syntax for denoting operator-overloading is to name a routine with the symbol of the operator and question marks where the arguments of the operation appear, e.g.:
153\begin{cfacode}
154int ++? (int op);                       //unary prefix increment
155int ?++ (int op);                       //unary postfix increment
156int ?+? (int op1, int op2);             //binary plus
157int ?<=?(int op1, int op2);             //binary less than
158int ?=? (int & op1, int op2);           //binary assignment
159int ?+=?(int & op1, int op2);           //binary plus-assignment
160
161struct S {int i, j;};
162S ?+?(S op1, S op2) {                           //add two structures
163        return (S){op1.i + op2.i, op1.j + op2.j};
164}
165S s1 = {1, 2}, s2 = {2, 3}, s3;
166s3 = s1 + s2;                                           //compute sum: s3 == {2, 5}
167\end{cfacode}
168While concurrency does not use operator overloading directly, this feature is more important as an introduction for the syntax of constructors.
169
170% ======================================================================
171\subsection{Constructors/Destructors}
172Object lifetime is often a challenge in concurrency. \CFA uses the approach of giving concurrent meaning to object lifetime as a means of synchronization and/or mutual exclusion. Since \CFA relies heavily on the lifetime of objects, constructors and destructors is a core feature required for concurrency and parallelism. \CFA uses the following syntax for constructors and destructors:
173\begin{cfacode}
174struct S {
175        size_t size;
176        int * ia;
177};
178void ?{}(S & s, int asize) {    //constructor operator
179        s.size = asize;                         //initialize fields
180        s.ia = calloc(size, sizeof(S));
181}
182void ^?{}(S & s) {                              //destructor operator
183        free(ia);                                       //de-initialization fields
184}
185int main() {
186        S x = {10}, y = {100};          //implicit calls: ?{}(x, 10), ?{}(y, 100)
187        ...                                                     //use x and y
188        ^x{}^y{};                            //explicit calls to de-initialize
189        x{20};  y{200};                         //explicit calls to reinitialize
190        ...                                                     //reuse x and y
191}                                                               //implicit calls: ^?{}(y), ^?{}(x)
192\end{cfacode}
193The language guarantees that every object and all their fields are constructed. Like \CC, construction of an object is automatically done on allocation and destruction of the object is done on deallocation. Allocation and deallocation can occur on the stack or on the heap.
194\begin{cfacode}
195{
196        struct S s = {10};      //allocation, call constructor
197        ...
198}                                               //deallocation, call destructor
199struct S * s = new();   //allocation, call constructor
200...
201delete(s);                              //deallocation, call destructor
202\end{cfacode}
203Note that like \CC, \CFA introduces \code{new} and \code{delete}, which behave like \code{malloc} and \code{free} in addition to constructing and destructing objects, after calling \code{malloc} and before calling \code{free}, respectively.
204
205% ======================================================================
206\subsection{Parametric Polymorphism}
207\label{s:ParametricPolymorphism}
208Routines in \CFA can also be reused for multiple types. This capability is done using the \code{forall} clauses, which allow separately compiled routines to support generic usage over multiple types. For example, the following sum function works for any type that supports construction from 0 and addition:
209\begin{cfacode}
210//constraint type, 0 and +
211forall(otype T | { void ?{}(T *, zero_t); T ?+?(T, T); })
212T sum(T a[ ], size_t size) {
213        T total = 0;                            //construct T from 0
214        for(size_t i = 0; i < size; i++)
215                total = total + a[i];   //select appropriate +
217}
218
219S sa[5];
220int i = sum(sa, 5);                             //use S's 0 construction and +
221\end{cfacode}
222
223Since writing constraints on types can become cumbersome for more constrained functions, \CFA also has the concept of traits. Traits are named collection of constraints that can be used both instead and in addition to regular constraints:
224\begin{cfacode}
225trait summable( otype T ) {
226        void ?{}(T *, zero_t);          //constructor from 0 literal
227        T ?+?(T, T);                            //assortment of additions
228        T ?+=?(T *, T);
229        T ++?(T *);
230        T ?++(T *);
231};
232forall( otype T | summable(T) ) //use trait
233T sum(T a[], size_t size);
234\end{cfacode}
235
236Note that the type use for assertions can be either an \code{otype} or a \code{dtype}. Types declared as \code{otype} refer to complete'' objects, i.e., objects with a size, a default constructor, a copy constructor, a destructor and an assignment operator. Using \code{dtype,} on the other hand, has none of these assumptions but is extremely restrictive, it only guarantees the object is addressable.
237
238% ======================================================================
239\subsection{with Clause/Statement}
240Since \CFA lacks the concept of a receiver, certain functions end up needing to repeat variable names often. To remove this inconvenience, \CFA provides the \code{with} statement, which opens an aggregate scope making its fields directly accessible (like Pascal).
241\begin{cfacode}
242struct S { int i, j; };
243int mem(S & this) with (this)           //with clause
244        i = 1;                                                  //this->i
245        j = 2;                                                  //this->j
246}
247int foo() {
248        struct S1 { ... } s1;
249        struct S2 { ... } s2;
250        with (s1)                                               //with statement
251        {
252                //access fields of s1 without qualification
253                with (s2)                                       //nesting
254                {
255                        //access fields of s1 and s2 without qualification
256                }
257        }
258        with (s1, s2)                                   //scopes open in parallel
259        {
260                //access fields of s1 and s2 without qualification
261        }
262}
263\end{cfacode}
264
266
267% ======================================================================
268% ======================================================================
269\section{Concurrency Basics}\label{basics}
270% ======================================================================
271% ======================================================================
272Before any detailed discussion of the concurrency and parallelism in \CFA, it is important to describe the basics of concurrency and how they are expressed in \CFA user code.
273
274\section{Basics of concurrency}
275At its core, concurrency is based on having multiple call-stacks and scheduling among threads of execution executing on these stacks. Concurrency without parallelism only requires having multiple call stacks (or contexts) for a single thread of execution.
276
277Execution with a single thread and multiple stacks where the thread is self-scheduling deterministically across the stacks is called coroutining. Execution with a single and multiple stacks but where the thread is scheduled by an oracle (non-deterministic from the thread's perspective) across the stacks is called concurrency.
278
279Therefore, a minimal concurrency system can be achieved by creating coroutines (see Section \ref{coroutine}), which instead of context-switching among each other, always ask an oracle where to context-switch next. While coroutines can execute on the caller's stack-frame, stack-full coroutines allow full generality and are sufficient as the basis for concurrency. The aforementioned oracle is a scheduler and the whole system now follows a cooperative threading-model (a.k.a., non-preemptive scheduling). The oracle/scheduler can either be a stack-less or stack-full entity and correspondingly require one or two context-switches to run a different coroutine. In any case, a subset of concurrency related challenges start to appear. For the complete set of concurrency challenges to occur, the only feature missing is preemption.
280
281A scheduler introduces order of execution uncertainty, while preemption introduces uncertainty about where context switches occur. Mutual exclusion and synchronization are ways of limiting non-determinism in a concurrent system. Now it is important to understand that uncertainty is desirable; uncertainty can be used by runtime systems to significantly increase performance and is often the basis of giving a user the illusion that tasks are running in parallel. Optimal performance in concurrent applications is often obtained by having as much non-determinism as correctness allows.
282
284One of the important features that are missing in C is threading\footnote{While the C11 standard defines a threads.h'' header, it is minimal and defined as optional. As such, library support for threading is far from widespread. At the time of writing the paper, neither \texttt{gcc} nor \texttt{clang} support threads.h'' in their respective standard libraries.}. On modern architectures, a lack of threading is unacceptable~\cite{Sutter05, Sutter05b}, and therefore modern programming languages must have the proper tools to allow users to write efficient concurrent programs to take advantage of parallelism. As an extension of C, \CFA needs to express these concepts in a way that is as natural as possible to programmers familiar with imperative languages. And being a system-level language means programmers expect to choose precisely which features they need and which cost they are willing to pay.
285
286\section{Coroutines: A Stepping Stone}\label{coroutine}
287While the main focus of this proposal is concurrency and parallelism, it is important to address coroutines, which are actually a significant building block of a concurrency system. \textbf{Coroutine}s are generalized routines which have predefined points where execution is suspended and can be resumed at a later time. Therefore, they need to deal with context switches and other context-management operations. This proposal includes coroutines both as an intermediate step for the implementation of threads, and a first-class feature of \CFA. Furthermore, many design challenges of threads are at least partially present in designing coroutines, which makes the design effort that much more relevant. The core \textbf{api} of coroutines revolves around two features: independent call-stacks and \code{suspend}/\code{resume}.
288
289\begin{table}
290\begin{center}
291\begin{tabular}{c @{\hskip 0.025in}|@{\hskip 0.025in} c @{\hskip 0.025in}|@{\hskip 0.025in} c}
292\begin{ccode}[tabsize=2]
293//Using callbacks
294void fibonacci_func(
295        int n,
296        void (*callback)(int)
297) {
298        int first = 0;
299        int second = 1;
300        int next, i;
301        for(i = 0; i < n; i++)
302        {
303                if(i <= 1)
304                        next = i;
305                else {
306                        next = f1 + f2;
307                        f1 = f2;
308                        f2 = next;
309                }
310                callback(next);
311        }
312}
313
314int main() {
315        void print_fib(int n) {
316                printf("%d\n", n);
317        }
318
319        fibonacci_func(
320                10, print_fib
321        );
322
323
324
325}
326\end{ccode}&\begin{ccode}[tabsize=2]
327//Using output array
328void fibonacci_array(
329        int n,
330        int* array
331) {
332        int f1 = 0; int f2 = 1;
333        int next, i;
334        for(i = 0; i < n; i++)
335        {
336                if(i <= 1)
337                        next = i;
338                else {
339                        next = f1 + f2;
340                        f1 = f2;
341                        f2 = next;
342                }
343                array[i] = next;
344        }
345}
346
347
348int main() {
349        int a[10];
350
351        fibonacci_func(
352                10, a
353        );
354
355        for(int i=0;i<10;i++){
356                printf("%d\n", a[i]);
357        }
358
359}
360\end{ccode}&\begin{ccode}[tabsize=2]
361//Using external state
362typedef struct {
363        int f1, f2;
364} Iterator_t;
365
366int fibonacci_state(
367        Iterator_t* it
368) {
369        int f;
370        f = it->f1 + it->f2;
371        it->f2 = it->f1;
372        it->f1 = max(f,1);
373        return f;
374}
375
376
377
378
379
380
381
382int main() {
383        Iterator_t it={0,0};
384
385        for(int i=0;i<10;i++){
386                printf("%d\n",
387                        fibonacci_state(
388                                &it
389                        );
390                );
391        }
392
393}
394\end{ccode}
395\end{tabular}
396\end{center}
397\caption{Different implementations of a Fibonacci sequence generator in C.}
398\label{lst:fibonacci-c}
399\end{table}
400
401A good example of a problem made easier with coroutines is generators, e.g., generating the Fibonacci sequence. This problem comes with the challenge of decoupling how a sequence is generated and how it is used. Listing \ref{lst:fibonacci-c} shows conventional approaches to writing generators in C. All three of these approach suffer from strong coupling. The left and centre approaches require that the generator have knowledge of how the sequence is used, while the rightmost approach requires holding internal state between calls on behalf of the generator and makes it much harder to handle corner cases like the Fibonacci seed.
402
403Listing \ref{lst:fibonacci-cfa} is an example of a solution to the Fibonacci problem using \CFA coroutines, where the coroutine stack holds sufficient state for the next generation. This solution has the advantage of having very strong decoupling between how the sequence is generated and how it is used. Indeed, this version is as easy to use as the \code{fibonacci_state} solution, while the implementation is very similar to the \code{fibonacci_func} example.
404
405\begin{figure}
406\begin{cfacode}[caption={Implementation of Fibonacci using coroutines},label={lst:fibonacci-cfa}]
407coroutine Fibonacci {
408        int fn; //used for communication
409};
410
411void ?{}(Fibonacci& this) { //constructor
412        this.fn = 0;
413}
414
415//main automatically called on first resume
416void main(Fibonacci& this) with (this) {
417        int fn1, fn2;           //retained between resumes
418        fn  = 0;
419        fn1 = fn;
421
422        fn  = 1;
423        fn2 = fn1;
424        fn1 = fn;
426
427        for ( ;; ) {
428                fn  = fn1 + fn2;
429                fn2 = fn1;
430                fn1 = fn;
432        }
433}
434
435int next(Fibonacci& this) {
436        resume(this); //transfer to last suspend
437        return this.fn;
438}
439
440void main() { //regular program main
441        Fibonacci f1, f2;
442        for ( int i = 1; i <= 10; i += 1 ) {
443                sout | next( f1 ) | next( f2 ) | endl;
444        }
445}
446\end{cfacode}
447\end{figure}
448
449Listing \ref{lst:fmt-line} shows the \code{Format} coroutine for restructuring text into groups of character blocks of fixed size. The example takes advantage of resuming coroutines in the constructor to simplify the code and highlights the idea that interesting control flow can occur in the constructor.
450
451\begin{figure}
452\begin{cfacode}[tabsize=3,caption={Formatting text into lines of 5 blocks of 4 characters.},label={lst:fmt-line}]
453//format characters into blocks of 4 and groups of 5 blocks per line
454coroutine Format {
455        char ch;                                                                        //used for communication
456        int g, b;                                                               //global because used in destructor
457};
458
459void  ?{}(Format& fmt) {
460        resume( fmt );                                                  //prime (start) coroutine
461}
462
463void ^?{}(Format& fmt) with fmt {
464        if ( fmt.g != 0 || fmt.b != 0 )
465        sout | endl;
466}
467
468void main(Format& fmt) with fmt {
469        for ( ;; ) {                                                    //for as many characters
470                for(g = 0; g < 5; g++) {                //groups of 5 blocks
471                        for(b = 0; b < 4; fb++) {       //blocks of 4 characters
472                                suspend();
473                                sout | ch;                                      //print character
474                        }
475                        sout | "  ";                                    //print block separator
476                }
477                sout | endl;                                            //print group separator
478        }
479}
480
481void prt(Format & fmt, char ch) {
482        fmt.ch = ch;
483        resume(fmt);
484}
485
486int main() {
487        Format fmt;
488        char ch;
489        Eof: for ( ;; ) {                                               //read until end of file
490                sin | ch;                                                       //read one character
491                if(eof(sin)) break Eof;                 //eof ?
492                prt(fmt, ch);                                           //push character for formatting
493        }
494}
495\end{cfacode}
496\end{figure}
497
498\subsection{Construction}
499One important design challenge for implementing coroutines and threads (shown in section \ref{threads}) is that the runtime system needs to run code after the user-constructor runs to connect the fully constructed object into the system. In the case of coroutines, this challenge is simpler since there is no non-determinism from preemption or scheduling. However, the underlying challenge remains the same for coroutines and threads.
500
501The runtime system needs to create the coroutine's stack and, more importantly, prepare it for the first resumption. The timing of the creation is non-trivial since users expect both to have fully constructed objects once execution enters the coroutine main and to be able to resume the coroutine from the constructor. There are several solutions to this problem but the chosen option effectively forces the design of the coroutine.
502
503Furthermore, \CFA faces an extra challenge as polymorphic routines create invisible thunks when cast to non-polymorphic routines and these thunks have function scope. For example, the following code, while looking benign, can run into undefined behaviour because of thunks:
504
505\begin{cfacode}
506//async: Runs function asynchronously on another thread
507forall(otype T)
508extern void async(void (*func)(T*), T* obj);
509
510forall(otype T)
511void noop(T*) {}
512
513void bar() {
514        int a;
515        async(noop, &a); //start thread running noop with argument a
516}
517\end{cfacode}
518
519The generated C code\footnote{Code trimmed down for brevity} creates a local thunk to hold type information:
520
521\begin{ccode}
522extern void async(/* omitted */, void (*func)(void*), void* obj);
523
524void noop(/* omitted */, void* obj){}
525
526void bar(){
527        int a;
528        void _thunk0(int* _p0){
529                /* omitted */
530                noop(/* omitted */, _p0);
531        }
532        /* omitted */
533        async(/* omitted */, ((void (*)(void*))(&_thunk0)), (&a));
534}
535\end{ccode}
536The problem in this example is a storage management issue, the function pointer \code{_thunk0} is only valid until the end of the block, which limits the viable solutions because storing the function pointer for too long causes undefined behaviour; i.e., the stack-based thunk being destroyed before it can be used. This challenge is an extension of challenges that come with second-class routines. Indeed, GCC nested routines also have the limitation that nested routine cannot be passed outside of the declaration scope. The case of coroutines and threads is simply an extension of this problem to multiple call stacks.
537
538\subsection{Alternative: Composition}
539One solution to this challenge is to use composition/containment, where coroutine fields are added to manage the coroutine.
540
541\begin{cfacode}
542struct Fibonacci {
543        int fn; //used for communication
544        coroutine c; //composition
545};
546
547void FibMain(void*) {
548        //...
549}
550
551void ?{}(Fibonacci& this) {
552        this.fn = 0;
553        //Call constructor to initialize coroutine
554        (this.c){myMain};
555}
556\end{cfacode}
557The downside of this approach is that users need to correctly construct the coroutine handle before using it. Like any other objects, the user must carefully choose construction order to prevent usage of objects not yet constructed. However, in the case of coroutines, users must also pass to the coroutine information about the coroutine main, like in the previous example. This opens the door for user errors and requires extra runtime storage to pass at runtime information that can be known statically.
558
559\subsection{Alternative: Reserved keyword}
560The next alternative is to use language support to annotate coroutines as follows:
561
562\begin{cfacode}
563coroutine Fibonacci {
564        int fn; //used for communication
565};
566\end{cfacode}
567The \code{coroutine} keyword means the compiler can find and inject code where needed. The downside of this approach is that it makes coroutine a special case in the language. Users wanting to extend coroutines or build their own for various reasons can only do so in ways offered by the language. Furthermore, implementing coroutines without language supports also displays the power of the programming language used. While this is ultimately the option used for idiomatic \CFA code, coroutines and threads can still be constructed by users without using the language support. The reserved keywords are only present to improve ease of use for the common cases.
568
569\subsection{Alternative: Lambda Objects}
570
571For coroutines as for threads, many implementations are based on routine pointers or function objects~\cite{Butenhof97, C++14, MS:VisualC++, BoostCoroutines15}. For example, Boost implements coroutines in terms of four functor object types:
572\begin{cfacode}
573asymmetric_coroutine<>::pull_type
574asymmetric_coroutine<>::push_type
575symmetric_coroutine<>::call_type
576symmetric_coroutine<>::yield_type
577\end{cfacode}
578Often, the canonical threading paradigm in languages is based on function pointers, \texttt{pthread} being one of the most well-known examples. The main problem of this approach is that the thread usage is limited to a generic handle that must otherwise be wrapped in a custom type. Since the custom type is simple to write in \CFA and solves several issues, added support for routine/lambda based coroutines adds very little.
579
580A variation of this would be to use a simple function pointer in the same way \texttt{pthread} does for threads:
581\begin{cfacode}
582void foo( coroutine_t cid, void* arg ) {
583        int* value = (int*)arg;
584        //Coroutine body
585}
586
587int main() {
588        int value = 0;
589        coroutine_t cid = coroutine_create( &foo, (void*)&value );
590        coroutine_resume( &cid );
591}
592\end{cfacode}
593This semantics is more common for thread interfaces but coroutines work equally well. As discussed in section \ref{threads}, this approach is superseded by static approaches in terms of expressivity.
594
595\subsection{Alternative: Trait-Based Coroutines}
596
597Finally, the underlying approach, which is the one closest to \CFA idioms, is to use trait-based lazy coroutines. This approach defines a coroutine as anything that satisfies the trait \code{is_coroutine} (as defined below) and is used as a coroutine.
598
599\begin{cfacode}
600trait is_coroutine(dtype T) {
601      void main(T& this);
602      coroutine_desc* get_coroutine(T& this);
603};
604
605forall( dtype T | is_coroutine(T) ) void suspend(T&);
606forall( dtype T | is_coroutine(T) ) void resume (T&);
607\end{cfacode}
608This ensures that an object is not a coroutine until \code{resume} is called on the object. Correspondingly, any object that is passed to \code{resume} is a coroutine since it must satisfy the \code{is_coroutine} trait to compile. The advantage of this approach is that users can easily create different types of coroutines, for example, changing the memory layout of a coroutine is trivial when implementing the \code{get_coroutine} routine. The \CFA keyword \code{coroutine} simply has the effect of implementing the getter and forward declarations required for users to implement the main routine.
609
610\begin{center}
611\begin{tabular}{c c c}
612\begin{cfacode}[tabsize=3]
613coroutine MyCoroutine {
614        int someValue;
615};
616\end{cfacode} & == & \begin{cfacode}[tabsize=3]
617struct MyCoroutine {
618        int someValue;
619        coroutine_desc __cor;
620};
621
622static inline
623coroutine_desc* get_coroutine(
624        struct MyCoroutine& this
625) {
626        return &this.__cor;
627}
628
629void main(struct MyCoroutine* this);
630\end{cfacode}
631\end{tabular}
632\end{center}
633
634The combination of these two approaches allows users new to coroutining and concurrency to have an easy and concise specification, while more advanced users have tighter control on memory layout and initialization.
635
637The basic building blocks of multithreading in \CFA are \textbf{cfathread}. Both user and kernel threads are supported, where user threads are the concurrency mechanism and kernel threads are the parallel mechanism. User threads offer a flexible and lightweight interface. A thread can be declared using a struct declaration \code{thread} as follows:
638
639\begin{cfacode}
641\end{cfacode}
642
643As for coroutines, the keyword is a thin wrapper around a \CFA trait:
644
645\begin{cfacode}
647      void ^?{}(T & mutex this);
648      void main(T & this);
650};
651\end{cfacode}
652
653Obviously, for this thread implementation to be useful it must run some user code. Several other threading interfaces use a function-pointer representation as the interface of threads (for example \Csharp~\cite{Csharp} and Scala~\cite{Scala}). However, this proposal considers that statically tying a \code{main} routine to a thread supersedes this approach. Since the \code{main} routine is already a special routine in \CFA (where the program begins), it is a natural extension of the semantics to use overloading to declare mains for different threads (the normal main being the main of the initial thread). As such the \code{main} routine of a thread can be defined as
654\begin{cfacode}
656
657void main(foo & this) {
658        sout | "Hello World!" | endl;
659}
660\end{cfacode}
661
662In this example, threads of type \code{foo} start execution in the \code{void main(foo &)} routine, which prints \code{"Hello World!".} While this paper encourages this approach to enforce strongly typed programming, users may prefer to use the routine-based thread semantics for the sake of simplicity. With the static semantics it is trivial to write a thread type that takes a function pointer as a parameter and executes it on its stack asynchronously.
663\begin{cfacode}
664typedef void (*voidFunc)(int);
665
667        voidFunc func;
668        int arg;
669};
670
671void ?{}(FuncRunner & this, voidFunc inFunc, int arg) {
672        this.func = inFunc;
673        this.arg  = arg;
674}
675
676void main(FuncRunner & this) {
677        //thread starts here and runs the function
678        this.func( this.arg );
679}
680
681void hello(/*unused*/ int) {
682        sout | "Hello World!" | endl;
683}
684
685int main() {
686        FuncRunner f = {hello, 42};
687        return 0?
688}
689\end{cfacode}
690
691A consequence of the strongly typed approach to main is that memory layout of parameters and return values to/from a thread are now explicitly specified in the \textbf{api}.
692
693Of course, for threads to be useful, it must be possible to start and stop threads and wait for them to complete execution. While using an \textbf{api} such as \code{fork} and \code{join} is relatively common in the literature, such an interface is unnecessary. Indeed, the simplest approach is to use \textbf{raii} principles and have threads \code{fork} after the constructor has completed and \code{join} before the destructor runs.
694\begin{cfacode}
696
697void main(World & this) {
698        sout | "World!" | endl;
699}
700
701void main() {
702        World w;
704
705        //Printing "Hello " and "World!" are run concurrently
706        sout | "Hello " | endl;
707
708        //Implicit join at end of scope
709}
710\end{cfacode}
711
712This semantic has several advantages over explicit semantics: a thread is always started and stopped exactly once, users cannot make any programming errors, and it naturally scales to multiple threads meaning basic synchronization is very simple.
713
714\begin{cfacode}
716        //...
717};
718
719//main
721        //...
722}
723
724void foo() {
726        //Start 10 threads at the beginning of the scope
727
728        DoStuff();
729
730        //Wait for the 10 threads to finish
731}
732\end{cfacode}
733
734However, one of the drawbacks of this approach is that threads always form a tree where nodes must always outlive their children, i.e., they are always destroyed in the opposite order of construction because of C scoping rules. This restriction is relaxed by using dynamic allocation, so threads can outlive the scope in which they are created, much like dynamically allocating memory lets objects outlive the scope in which they are created.
735
736\begin{cfacode}
738        //...
739};
740
742        //...
743}
744
745void foo() {
747        {
748                //Start a thread at the beginning of the scope
750
751                //create another thread that will outlive the thread in this scope
753
754                DoStuff();
755
756                //Wait for the thread short_lived to finish
757        }
758        DoMoreStuff();
759
760        //Now wait for the long_lived to finish
761        delete long_lived;
762}
763\end{cfacode}
764
765
766% ======================================================================
767% ======================================================================
768\section{Concurrency}
769% ======================================================================
770% ======================================================================
771Several tools can be used to solve concurrency challenges. Since many of these challenges appear with the use of mutable shared state, some languages and libraries simply disallow mutable shared state (Erlang~\cite{Erlang}, Haskell~\cite{Haskell}, Akka (Scala)~\cite{Akka}). In these paradigms, interaction among concurrent objects relies on message passing~\cite{Thoth,Harmony,V-Kernel} or other paradigms closely relate to networking concepts (channels~\cite{CSP,Go} for example). However, in languages that use routine calls as their core abstraction mechanism, these approaches force a clear distinction between concurrent and non-concurrent paradigms (i.e., message passing versus routine calls). This distinction in turn means that, in order to be effective, programmers need to learn two sets of design patterns. While this distinction can be hidden away in library code, effective use of the library still has to take both paradigms into account.
772
773Approaches based on shared memory are more closely related to non-concurrent paradigms since they often rely on basic constructs like routine calls and shared objects. At the lowest level, concurrent paradigms are implemented as atomic operations and locks. Many such mechanisms have been proposed, including semaphores~\cite{Dijkstra68b} and path expressions~\cite{Campbell74}. However, for productivity reasons it is desirable to have a higher-level construct be the core concurrency paradigm~\cite{HPP:Study}.
774
775An approach that is worth mentioning because it is gaining in popularity is transactional memory~\cite{Herlihy93}. While this approach is even pursued by system languages like \CC~\cite{Cpp-Transactions}, the performance and feature set is currently too restrictive to be the main concurrency paradigm for system languages, which is why it was rejected as the core paradigm for concurrency in \CFA.
776
777One of the most natural, elegant, and efficient mechanisms for synchronization and communication, especially for shared-memory systems, is the \emph{monitor}. Monitors were first proposed by Brinch Hansen~\cite{Hansen73} and later described and extended by C.A.R.~Hoare~\cite{Hoare74}. Many programming languages---e.g., Concurrent Pascal~\cite{ConcurrentPascal}, Mesa~\cite{Mesa}, Modula~\cite{Modula-2}, Turing~\cite{Turing:old}, Modula-3~\cite{Modula-3}, NeWS~\cite{NeWS}, Emerald~\cite{Emerald}, \uC~\cite{Buhr92a} and Java~\cite{Java}---provide monitors as explicit language constructs. In addition, operating-system kernels and device drivers have a monitor-like structure, although they often use lower-level primitives such as semaphores or locks to simulate monitors. For these reasons, this project proposes monitors as the core concurrency construct.
778
779\section{Basics}
780Non-determinism requires concurrent systems to offer support for mutual-exclusion and synchronization. Mutual-exclusion is the concept that only a fixed number of threads can access a critical section at any given time, where a critical section is a group of instructions on an associated portion of data that requires the restricted access. On the other hand, synchronization enforces relative ordering of execution and synchronization tools provide numerous mechanisms to establish timing relationships among threads.
781
782\subsection{Mutual-Exclusion}
783As mentioned above, mutual-exclusion is the guarantee that only a fix number of threads can enter a critical section at once. However, many solutions exist for mutual exclusion, which vary in terms of performance, flexibility and ease of use. Methods range from low-level locks, which are fast and flexible but require significant attention to be correct, to  higher-level concurrency techniques, which sacrifice some performance in order to improve ease of use. Ease of use comes by either guaranteeing some problems cannot occur (e.g., being deadlock free) or by offering a more explicit coupling between data and corresponding critical section. For example, the \CC \code{std::atomic<T>} offers an easy way to express mutual-exclusion on a restricted set of operations (e.g., reading/writing large types atomically). Another challenge with low-level locks is composability. Locks have restricted composability because it takes careful organizing for multiple locks to be used while preventing deadlocks. Easing composability is another feature higher-level mutual-exclusion mechanisms often offer.
784
785\subsection{Synchronization}
786As with mutual-exclusion, low-level synchronization primitives often offer good performance and good flexibility at the cost of ease of use. Again, higher-level mechanisms often simplify usage by adding either better coupling between synchronization and data (e.g., message passing) or offering a simpler solution to otherwise involved challenges. As mentioned above, synchronization can be expressed as guaranteeing that event \textit{X} always happens before \textit{Y}. Most of the time, synchronization happens within a critical section, where threads must acquire mutual-exclusion in a certain order. However, it may also be desirable to guarantee that event \textit{Z} does not occur between \textit{X} and \textit{Y}. Not satisfying this property is called \textbf{barging}. For example, where event \textit{X} tries to effect event \textit{Y} but another thread acquires the critical section and emits \textit{Z} before \textit{Y}. The classic example is the thread that finishes using a resource and unblocks a thread waiting to use the resource, but the unblocked thread must compete to acquire the resource. Preventing or detecting barging is an involved challenge with low-level locks, which can be made much easier by higher-level constructs. This challenge is often split into two different methods, barging avoidance and barging prevention. Algorithms that use flag variables to detect barging threads are said to be using barging avoidance, while algorithms that baton-pass locks~\cite{Andrews89} between threads instead of releasing the locks are said to be using barging prevention.
787
788% ======================================================================
789% ======================================================================
790\section{Monitors}
791% ======================================================================
792% ======================================================================
793A \textbf{monitor} is a set of routines that ensure mutual-exclusion when accessing shared state. More precisely, a monitor is a programming technique that associates mutual-exclusion to routine scopes, as opposed to mutex locks, where mutual-exclusion is defined by lock/release calls independently of any scoping of the calling routine. This strong association eases readability and maintainability, at the cost of flexibility. Note that both monitors and mutex locks, require an abstract handle to identify them. This concept is generally associated with object-oriented languages like Java~\cite{Java} or \uC~\cite{uC++book} but does not strictly require OO semantics. The only requirement is the ability to declare a handle to a shared object and a set of routines that act on it:
794\begin{cfacode}
795typedef /*some monitor type*/ monitor;
796int f(monitor & m);
797
798int main() {
799        monitor m;  //Handle m
800        f(m);       //Routine using handle
801}
802\end{cfacode}
803
804% ======================================================================
805% ======================================================================
806\subsection{Call Semantics} \label{call}
807% ======================================================================
808% ======================================================================
809The above monitor example displays some of the intrinsic characteristics. First, it is necessary to use pass-by-reference over pass-by-value for monitor routines. This semantics is important, because at their core, monitors are implicit mutual-exclusion objects (locks), and these objects cannot be copied. Therefore, monitors are non-copy-able objects (\code{dtype}).
810
811Another aspect to consider is when a monitor acquires its mutual exclusion. For example, a monitor may need to be passed through multiple helper routines that do not acquire the monitor mutual-exclusion on entry. Passthrough can occur for generic helper routines (\code{swap}, \code{sort}, etc.) or specific helper routines like the following to implement an atomic counter:
812
813\begin{cfacode}
814monitor counter_t { /*...see section $\ref{data}$...*/ };
815
816void ?{}(counter_t & nomutex this); //constructor
817size_t ++?(counter_t & mutex this); //increment
818
819//need for mutex is platform dependent
820void ?{}(size_t * this, counter_t & mutex cnt); //conversion
821\end{cfacode}
822This counter is used as follows:
823\begin{center}
824\begin{tabular}{c @{\hskip 0.35in} c @{\hskip 0.35in} c}
825\begin{cfacode}
826//shared counter
827counter_t cnt1, cnt2;
828
833        ...
835\end{cfacode}
836\end{tabular}
837\end{center}
838Notice how the counter is used without any explicit synchronization and yet supports thread-safe semantics for both reading and writing, which is similar in usage to the \CC template \code{std::atomic}.
839
840Here, the constructor (\code{?\{\}}) uses the \code{nomutex} keyword to signify that it does not acquire the monitor mutual-exclusion when constructing. This semantics is because an object not yet con\-structed should never be shared and therefore does not require mutual exclusion. Furthermore, it allows the implementation greater freedom when it initializes the monitor locking. The prefix increment operator uses \code{mutex} to protect the incrementing process from race conditions. Finally, there is a conversion operator from \code{counter_t} to \code{size_t}. This conversion may or may not require the \code{mutex} keyword depending on whether or not reading a \code{size_t} is an atomic operation.
841
842For maximum usability, monitors use \textbf{multi-acq} semantics, which means a single thread can acquire the same monitor multiple times without deadlock. For example, listing \ref{fig:search} uses recursion and \textbf{multi-acq} to print values inside a binary tree.
843\begin{figure}
844\begin{cfacode}[caption={Recursive printing algorithm using \textbf{multi-acq}.},label={fig:search}]
845monitor printer { ... };
846struct tree {
847        tree * left, right;
848        char * value;
849};
850void print(printer & mutex p, char * v);
851
852void print(printer & mutex p, tree * t) {
853        print(p, t->value);
854        print(p, t->left );
855        print(p, t->right);
856}
857\end{cfacode}
858\end{figure}
859
860Having both \code{mutex} and \code{nomutex} keywords can be redundant, depending on the meaning of a routine having neither of these keywords. For example, it is reasonable that it should default to the safest option (\code{mutex}) when given a routine without qualifiers \code{void foo(counter_t & this)}, whereas assuming \code{nomutex} is unsafe and may cause subtle errors. On the other hand, \code{nomutex} is the normal'' parameter behaviour, it effectively states explicitly that this routine is not special''. Another alternative is making exactly one of these keywords mandatory, which provides the same semantics but without the ambiguity of supporting routines with neither keyword. Mandatory keywords would also have the added benefit of being self-documented but at the cost of extra typing. While there are several benefits to mandatory keywords, they do bring a few challenges. Mandatory keywords in \CFA would imply that the compiler must know without doubt whether or not a parameter is a monitor or not. Since \CFA relies heavily on traits as an abstraction mechanism, the distinction between a type that is a monitor and a type that looks like a monitor can become blurred. For this reason, \CFA only has the \code{mutex} keyword and uses no keyword to mean \code{nomutex}.
861
862The next semantic decision is to establish when \code{mutex} may be used as a type qualifier. Consider the following declarations:
863\begin{cfacode}
864int f1(monitor & mutex m);
865int f2(const monitor & mutex m);
866int f3(monitor ** mutex m);
867int f4(monitor * mutex m []);
868int f5(graph(monitor *) & mutex m);
869\end{cfacode}
870The problem is to identify which object(s) should be acquired. Furthermore, each object needs to be acquired only once. In the case of simple routines like \code{f1} and \code{f2} it is easy to identify an exhaustive list of objects to acquire on entry. Adding indirections (\code{f3}) still allows the compiler and programmer to identify which object is acquired. However, adding in arrays (\code{f4}) makes it much harder. Array lengths are not necessarily known in C, and even then, making sure objects are only acquired once becomes none-trivial. This problem can be extended to absurd limits like \code{f5}, which uses a graph of monitors. To make the issue tractable, this project imposes the requirement that a routine may only acquire one monitor per parameter and it must be the type of the parameter with at most one level of indirection (ignoring potential qualifiers). Also note that while routine \code{f3} can be supported, meaning that monitor \code{**m} is acquired, passing an array to this routine would be type-safe and yet result in undefined behaviour because only the first element of the array is acquired. However, this ambiguity is part of the C type-system with respects to arrays. For this reason, \code{mutex} is disallowed in the context where arrays may be passed:
871\begin{cfacode}
872int f1(monitor & mutex m);    //Okay : recommended case
873int f2(monitor * mutex m);    //Not Okay : Could be an array
874int f3(monitor mutex m []);  //Not Okay : Array of unknown length
875int f4(monitor ** mutex m);   //Not Okay : Could be an array
876int f5(monitor * mutex m []); //Not Okay : Array of unknown length
877\end{cfacode}
878Note that not all array functions are actually distinct in the type system. However, even if the code generation could tell the difference, the extra information is still not sufficient to extend meaningfully the monitor call semantic.
879
880Unlike object-oriented monitors, where calling a mutex member \emph{implicitly} acquires mutual-exclusion of the receiver object, \CFA uses an explicit mechanism to specify the object that acquires mutual-exclusion. A consequence of this approach is that it extends naturally to multi-monitor calls.
881\begin{cfacode}
882int f(MonitorA & mutex a, MonitorB & mutex b);
883
884MonitorA a;
885MonitorB b;
886f(a,b);
887\end{cfacode}
888While OO monitors could be extended with a mutex qualifier for multiple-monitor calls, no example of this feature could be found. The capability to acquire multiple locks before entering a critical section is called \emph{\textbf{bulk-acq}}. In practice, writing multi-locking routines that do not lead to deadlocks is tricky. Having language support for such a feature is therefore a significant asset for \CFA. In the case presented above, \CFA guarantees that the order of acquisition is consistent across calls to different routines using the same monitors as arguments. This consistent ordering means acquiring multiple monitors is safe from deadlock when using \textbf{bulk-acq}. However, users can still force the acquiring order. For example, notice which routines use \code{mutex}/\code{nomutex} and how this affects acquiring order:
889\begin{cfacode}
890void foo(A& mutex a, B& mutex b) { //acquire a & b
891        ...
892}
893
894void bar(A& mutex a, B& /*nomutex*/ b) { //acquire a
895        ... foo(a, b); ... //acquire b
896}
897
898void baz(A& /*nomutex*/ a, B& mutex b) { //acquire b
899        ... foo(a, b); ... //acquire a
900}
901\end{cfacode}
902The \textbf{multi-acq} monitor lock allows a monitor lock to be acquired by both \code{bar} or \code{baz} and acquired again in \code{foo}. In the calls to \code{bar} and \code{baz} the monitors are acquired in opposite order.
903
904However, such use leads to lock acquiring order problems. In the example above, the user uses implicit ordering in the case of function \code{foo} but explicit ordering in the case of \code{bar} and \code{baz}. This subtle difference means that calling these routines concurrently may lead to deadlock and is therefore undefined behaviour. As shown~\cite{Lister77}, solving this problem requires:
905\begin{enumerate}
906        \item Dynamically tracking the monitor-call order.
907        \item Implement rollback semantics.
908\end{enumerate}
909While the first requirement is already a significant constraint on the system, implementing a general rollback semantics in a C-like language is still prohibitively complex~\cite{Dice10}. In \CFA, users simply need to be careful when acquiring multiple monitors at the same time or only use \textbf{bulk-acq} of all the monitors. While \CFA provides only a partial solution, most systems provide no solution and the \CFA partial solution handles many useful cases.
910
911For example, \textbf{multi-acq} and \textbf{bulk-acq} can be used together in interesting ways:
912\begin{cfacode}
913monitor bank { ... };
914
915void deposit( bank & mutex b, int deposit );
916
917void transfer( bank & mutex mybank, bank & mutex yourbank, int me2you) {
918        deposit( mybank, -me2you );
919        deposit( yourbank, me2you );
920}
921\end{cfacode}
922This example shows a trivial solution to the bank-account transfer problem~\cite{BankTransfer}. Without \textbf{multi-acq} and \textbf{bulk-acq}, the solution to this problem is much more involved and requires careful engineering.
923
924\subsection{\code{mutex} statement} \label{mutex-stmt}
925
926The call semantics discussed above have one software engineering issue: only a routine can acquire the mutual-exclusion of a set of monitor. \CFA offers the \code{mutex} statement to work around the need for unnecessary names, avoiding a major software engineering problem~\cite{2FTwoHardThings}. Table \ref{lst:mutex-stmt} shows an example of the \code{mutex} statement, which introduces a new scope in which the mutual-exclusion of a set of monitor is acquired. Beyond naming, the \code{mutex} statement has no semantic difference from a routine call with \code{mutex} parameters.
927
928\begin{table}
929\begin{center}
930\begin{tabular}{|c|c|}
931function call & \code{mutex} statement \\
932\hline
933\begin{cfacode}[tabsize=3]
934monitor M {};
935void foo( M & mutex m1, M & mutex m2 ) {
936        //critical section
937}
938
939void bar( M & m1, M & m2 ) {
940        foo( m1, m2 );
941}
942\end{cfacode}&\begin{cfacode}[tabsize=3]
943monitor M {};
944void bar( M & m1, M & m2 ) {
945        mutex(m1, m2) {
946                //critical section
947        }
948}
949
950
951\end{cfacode}
952\end{tabular}
953\end{center}
954\caption{Regular call semantics vs. \code{mutex} statement}
955\label{lst:mutex-stmt}
956\end{table}
957
958% ======================================================================
959% ======================================================================
960\subsection{Data semantics} \label{data}
961% ======================================================================
962% ======================================================================
963Once the call semantics are established, the next step is to establish data semantics. Indeed, until now a monitor is used simply as a generic handle but in most cases monitors contain shared data. This data should be intrinsic to the monitor declaration to prevent any accidental use of data without its appropriate protection. For example, here is a complete version of the counter shown in section \ref{call}:
964\begin{cfacode}
965monitor counter_t {
966        int value;
967};
968
969void ?{}(counter_t & this) {
970        this.cnt = 0;
971}
972
973int ?++(counter_t & mutex this) {
974        return ++this.value;
975}
976
977//need for mutex is platform dependent here
978void ?{}(int * this, counter_t & mutex cnt) {
979        *this = (int)cnt;
980}
981\end{cfacode}
982
983Like threads and coroutines, monitors are defined in terms of traits with some additional language support in the form of the \code{monitor} keyword. The monitor trait is:
984\begin{cfacode}
985trait is_monitor(dtype T) {
986        monitor_desc * get_monitor( T & );
987        void ^?{}( T & mutex );
988};
989\end{cfacode}
990Note that the destructor of a monitor must be a \code{mutex} routine to prevent deallocation while a thread is accessing the monitor. As with any object, calls to a monitor, using \code{mutex} or otherwise, is undefined behaviour after the destructor has run.
991
992% ======================================================================
993% ======================================================================
994\section{Internal Scheduling} \label{intsched}
995% ======================================================================
996% ======================================================================
997In addition to mutual exclusion, the monitors at the core of \CFA's concurrency can also be used to achieve synchronization. With monitors, this capability is generally achieved with internal or external scheduling as in~\cite{Hoare74}. With \textbf{scheduling} loosely defined as deciding which thread acquires the critical section next, \textbf{internal scheduling} means making the decision from inside the critical section (i.e., with access to the shared state), while \textbf{external scheduling} means making the decision when entering the critical section (i.e., without access to the shared state). Since internal scheduling within a single monitor is mostly a solved problem, this paper concentrates on extending internal scheduling to multiple monitors. Indeed, like the \textbf{bulk-acq} semantics, internal scheduling extends to multiple monitors in a way that is natural to the user but requires additional complexity on the implementation side.
998
999First, here is a simple example of internal scheduling:
1000
1001\begin{cfacode}
1002monitor A {
1003        condition e;
1004}
1005
1006void foo(A& mutex a1, A& mutex a2) {
1007        ...
1008        //Wait for cooperation from bar()
1009        wait(a1.e);
1010        ...
1011}
1012
1013void bar(A& mutex a1, A& mutex a2) {
1014        //Provide cooperation for foo()
1015        ...
1016        //Unblock foo
1017        signal(a1.e);
1018}
1019\end{cfacode}
1020There are two details to note here. First, \code{signal} is a delayed operation; it only unblocks the waiting thread when it reaches the end of the critical section. This semantics is needed to respect mutual-exclusion, i.e., the signaller and signalled thread cannot be in the monitor simultaneously. The alternative is to return immediately after the call to \code{signal}, which is significantly more restrictive. Second, in \CFA, while it is common to store a \code{condition} as a field of the monitor, a \code{condition} variable can be stored/created independently of a monitor. Here routine \code{foo} waits for the \code{signal} from \code{bar} before making further progress, ensuring a basic ordering.
1021
1022An important aspect of the implementation is that \CFA does not allow barging, which means that once function \code{bar} releases the monitor, \code{foo} is guaranteed to be the next thread to acquire the monitor (unless some other thread waited on the same condition). This guarantee offers the benefit of not having to loop around waits to recheck that a condition is met. The main reason \CFA offers this guarantee is that users can easily introduce barging if it becomes a necessity but adding barging prevention or barging avoidance is more involved without language support. Supporting barging prevention as well as extending internal scheduling to multiple monitors is the main source of complexity in the design and implementation of \CFA concurrency.
1023
1024% ======================================================================
1025% ======================================================================
1026\subsection{Internal Scheduling - Multi-Monitor}
1027% ======================================================================
1028% ======================================================================
1029It is easy to understand the problem of multi-monitor scheduling using a series of pseudo-code examples. Note that for simplicity in the following snippets of pseudo-code, waiting and signalling is done using an implicit condition variable, like Java built-in monitors. Indeed, \code{wait} statements always use the implicit condition variable as parameters and explicitly name the monitors (A and B) associated with the condition. Note that in \CFA, condition variables are tied to a \emph{group} of monitors on first use (called branding), which means that using internal scheduling with distinct sets of monitors requires one condition variable per set of monitors. The example below shows the simple case of having two threads (one for each column) and a single monitor A.
1030
1031\begin{multicols}{2}
1033\begin{pseudo}
1034acquire A
1035        wait A
1036release A
1037\end{pseudo}
1038
1039\columnbreak
1040
1042\begin{pseudo}
1043acquire A
1044        signal A
1045release A
1046\end{pseudo}
1047\end{multicols}
1048One thread acquires before waiting (atomically blocking and releasing A) and the other acquires before signalling. It is important to note here that both \code{wait} and \code{signal} must be called with the proper monitor(s) already acquired. This semantic is a logical requirement for barging prevention.
1049
1050A direct extension of the previous example is a \textbf{bulk-acq} version:
1051\begin{multicols}{2}
1052\begin{pseudo}
1053acquire A & B
1054        wait A & B
1055release A & B
1056\end{pseudo}
1057\columnbreak
1058\begin{pseudo}
1059acquire A & B
1060        signal A & B
1061release A & B
1062\end{pseudo}
1063\end{multicols}
1064\noindent This version uses \textbf{bulk-acq} (denoted using the {\sf\&} symbol), but the presence of multiple monitors does not add a particularly new meaning. Synchronization happens between the two threads in exactly the same way and order. The only difference is that mutual exclusion covers a group of monitors. On the implementation side, handling multiple monitors does add a degree of complexity as the next few examples demonstrate.
1065
1066While deadlock issues can occur when nesting monitors, these issues are only a symptom of the fact that locks, and by extension monitors, are not perfectly composable. For monitors, a well-known deadlock problem is the Nested Monitor Problem~\cite{Lister77}, which occurs when a \code{wait} is made by a thread that holds more than one monitor. For example, the following pseudo-code runs into the nested-monitor problem:
1067\begin{multicols}{2}
1068\begin{pseudo}
1069acquire A
1070        acquire B
1071                wait B
1072        release B
1073release A
1074\end{pseudo}
1075
1076\columnbreak
1077
1078\begin{pseudo}
1079acquire A
1080        acquire B
1081                signal B
1082        release B
1083release A
1084\end{pseudo}
1085\end{multicols}
1086\noindent The \code{wait} only releases monitor \code{B} so the signalling thread cannot acquire monitor \code{A} to get to the \code{signal}. Attempting release of all acquired monitors at the \code{wait} introduces a different set of problems, such as releasing monitor \code{C}, which has nothing to do with the \code{signal}.
1087
1088However, for monitors as for locks, it is possible to write a program using nesting without encountering any problems if nesting is done correctly. For example, the next pseudo-code snippet acquires monitors {\sf A} then {\sf B} before waiting, while only acquiring {\sf B} when signalling, effectively avoiding the Nested Monitor Problem~\cite{Lister77}.
1089
1090\begin{multicols}{2}
1091\begin{pseudo}
1092acquire A
1093        acquire B
1094                wait B
1095        release B
1096release A
1097\end{pseudo}
1098
1099\columnbreak
1100
1101\begin{pseudo}
1102
1103acquire B
1104        signal B
1105release B
1106
1107\end{pseudo}
1108\end{multicols}
1109
1110\noindent However, this simple refactoring may not be possible, forcing more complex restructuring.
1111
1112% ======================================================================
1113% ======================================================================
1114\subsection{Internal Scheduling - In Depth}
1115% ======================================================================
1116% ======================================================================
1117
1118A larger example is presented to show complex issues for \textbf{bulk-acq} and its implementation options are analyzed. Listing \ref{lst:int-bulk-pseudo} shows an example where \textbf{bulk-acq} adds a significant layer of complexity to the internal signalling semantics, and listing \ref{lst:int-bulk-cfa} shows the corresponding \CFA code to implement the pseudo-code in listing \ref{lst:int-bulk-pseudo}. For the purpose of translating the given pseudo-code into \CFA-code, any method of introducing a monitor is acceptable, e.g., \code{mutex} parameters, global variables, pointer parameters, or using locals with the \code{mutex} statement.
1119
1120\begin{figure}[!t]
1121\begin{multicols}{2}
1123\begin{pseudo}[numbers=left]
1124acquire A
1125        //Code Section 1
1126        acquire A & B
1127                //Code Section 2
1128                wait A & B
1129                //Code Section 3
1130        release A & B
1131        //Code Section 4
1132release A
1133\end{pseudo}
1134\columnbreak
1136\begin{pseudo}[numbers=left, firstnumber=10,escapechar=|]
1137acquire A
1138        //Code Section 5
1139        acquire A & B
1140                //Code Section 6
1141                |\label{line:signal1}|signal A & B
1142                //Code Section 7
1143        |\label{line:releaseFirst}|release A & B
1144        //Code Section 8
1145|\label{line:lastRelease}|release A
1146\end{pseudo}
1147\end{multicols}
1148\begin{cfacode}[caption={Internal scheduling with \textbf{bulk-acq}},label={lst:int-bulk-pseudo}]
1149\end{cfacode}
1150\begin{center}
1151\begin{cfacode}[xleftmargin=.4\textwidth]
1152monitor A a;
1153monitor B b;
1154condition c;
1155\end{cfacode}
1156\end{center}
1157\begin{multicols}{2}
1159\begin{cfacode}
1160mutex(a) {
1161        //Code Section 1
1162        mutex(a, b) {
1163                //Code Section 2
1164                wait(c);
1165                //Code Section 3
1166        }
1167        //Code Section 4
1168}
1169\end{cfacode}
1170\columnbreak
1172\begin{cfacode}
1173mutex(a) {
1174        //Code Section 5
1175        mutex(a, b) {
1176                //Code Section 6
1177                signal(c);
1178                //Code Section 7
1179        }
1180        //Code Section 8
1181}
1182\end{cfacode}
1183\end{multicols}
1184\begin{cfacode}[caption={Equivalent \CFA code for listing \ref{lst:int-bulk-pseudo}},label={lst:int-bulk-cfa}]
1185\end{cfacode}
1186\begin{multicols}{2}
1187Waiter
1188\begin{pseudo}[numbers=left]
1189acquire A
1190        acquire A & B
1191                wait A & B
1192        release A & B
1193release A
1194\end{pseudo}
1195
1196\columnbreak
1197
1198Signaller
1199\begin{pseudo}[numbers=left, firstnumber=6,escapechar=|]
1200acquire A
1201        acquire A & B
1202                signal A & B
1203        release A & B
1204        |\label{line:secret}|//Secretly keep B here
1205release A
1206//Wakeup waiter and transfer A & B
1207\end{pseudo}
1208\end{multicols}
1209\begin{cfacode}[caption={Listing \ref{lst:int-bulk-pseudo}, with delayed signalling comments},label={lst:int-secret}]
1210\end{cfacode}
1211\end{figure}
1212
1213The complexity begins at code sections 4 and 8 in listing \ref{lst:int-bulk-pseudo}, which are where the existing semantics of internal scheduling needs to be extended for multiple monitors. The root of the problem is that \textbf{bulk-acq} is used in a context where one of the monitors is already acquired, which is why it is important to define the behaviour of the previous pseudo-code. When the signaller thread reaches the location where it should release \code{A & B}'' (listing \ref{lst:int-bulk-pseudo} line \ref{line:releaseFirst}), it must actually transfer ownership of monitor \code{B} to the waiting thread. This ownership transfer is required in order to prevent barging into \code{B} by another thread, since both the signalling and signalled threads still need monitor \code{A}. There are three options:
1214
1215\subsubsection{Delaying Signals}
1216The obvious solution to the problem of multi-monitor scheduling is to keep ownership of all locks until the last lock is ready to be transferred. It can be argued that that moment is when the last lock is no longer needed, because this semantics fits most closely to the behaviour of single-monitor scheduling. This solution has the main benefit of transferring ownership of groups of monitors, which simplifies the semantics from multiple objects to a single group of objects, effectively making the existing single-monitor semantic viable by simply changing monitors to monitor groups. This solution releases the monitors once every monitor in a group can be released. However, since some monitors are never released (e.g., the monitor of a thread), this interpretation means a group might never be released. A more interesting interpretation is to transfer the group until all its monitors are released, which means the group is not passed further and a thread can retain its locks.
1217
1218However, listing \ref{lst:int-secret} shows this solution can become much more complicated depending on what is executed while secretly holding B at line \ref{line:secret}, while avoiding the need to transfer ownership of a subset of the condition monitors. Listing \ref{lst:dependency} shows a slightly different example where a third thread is waiting on monitor \code{A}, using a different condition variable. Because the third thread is signalled when secretly holding \code{B}, the goal  becomes unreachable. Depending on the order of signals (listing \ref{lst:dependency} line \ref{line:signal-ab} and \ref{line:signal-a}) two cases can happen:
1219
1220\paragraph{Case 1: thread $\alpha$ goes first.} In this case, the problem is that monitor \code{A} needs to be passed to thread $\beta$ when thread $\alpha$ is done with it.
1221\paragraph{Case 2: thread $\beta$ goes first.} In this case, the problem is that monitor \code{B} needs to be retained and passed to thread $\alpha$ along with monitor \code{A}, which can be done directly or possibly using thread $\beta$ as an intermediate.
1222\\
1223
1224Note that ordering is not determined by a race condition but by whether signalled threads are enqueued in FIFO or FILO order. However, regardless of the answer, users can move line \ref{line:signal-a} before line \ref{line:signal-ab} and get the reverse effect for listing \ref{lst:dependency}.
1225
1226In both cases, the threads need to be able to distinguish, on a per monitor basis, which ones need to be released and which ones need to be transferred, which means knowing when to release a group becomes complex and inefficient (see next section) and therefore effectively precludes this approach.
1227
1228\subsubsection{Dependency graphs}
1229
1230
1231\begin{figure}
1232\begin{multicols}{3}
1233Thread $\alpha$
1234\begin{pseudo}[numbers=left, firstnumber=1]
1235acquire A
1236        acquire A & B
1237                wait A & B
1238        release A & B
1239release A
1240\end{pseudo}
1241\columnbreak
1242Thread $\gamma$
1243\begin{pseudo}[numbers=left, firstnumber=6, escapechar=|]
1244acquire A
1245        acquire A & B
1246                |\label{line:signal-ab}|signal A & B
1247        |\label{line:release-ab}|release A & B
1248        |\label{line:signal-a}|signal A
1249|\label{line:release-a}|release A
1250\end{pseudo}
1251\columnbreak
1252Thread $\beta$
1253\begin{pseudo}[numbers=left, firstnumber=12, escapechar=|]
1254acquire A
1255        wait A
1256|\label{line:release-aa}|release A
1257\end{pseudo}
1258\end{multicols}
1259\begin{cfacode}[caption={Pseudo-code for the three thread example.},label={lst:dependency}]
1260\end{cfacode}
1261\begin{center}
1262\input{dependency}
1263\end{center}
1264\caption{Dependency graph of the statements in listing \ref{lst:dependency}}
1265\label{fig:dependency}
1266\end{figure}
1267
1268In listing \ref{lst:int-bulk-pseudo}, there is a solution that satisfies both barging prevention and mutual exclusion. If ownership of both monitors is transferred to the waiter when the signaller releases \code{A & B} and then the waiter transfers back ownership of \code{A} back to the signaller when it releases it, then the problem is solved (\code{B} is no longer in use at this point). Dynamically finding the correct order is therefore the second possible solution. The problem is effectively resolving a dependency graph of ownership requirements. Here even the simplest of code snippets requires two transfers and has a super-linear complexity. This complexity can be seen in listing \ref{lst:explosion}, which is just a direct extension to three monitors, requires at least three ownership transfer and has multiple solutions. Furthermore, the presence of multiple solutions for ownership transfer can cause deadlock problems if a specific solution is not consistently picked; In the same way that multiple lock acquiring order can cause deadlocks.
1269\begin{figure}
1270\begin{multicols}{2}
1271\begin{pseudo}
1272acquire A
1273        acquire B
1274                acquire C
1275                        wait A & B & C
1276                release C
1277        release B
1278release A
1279\end{pseudo}
1280
1281\columnbreak
1282
1283\begin{pseudo}
1284acquire A
1285        acquire B
1286                acquire C
1287                        signal A & B & C
1288                release C
1289        release B
1290release A
1291\end{pseudo}
1292\end{multicols}
1293\begin{cfacode}[caption={Extension to three monitors of listing \ref{lst:int-bulk-pseudo}},label={lst:explosion}]
1294\end{cfacode}
1295\end{figure}
1296
1297Given the three threads example in listing \ref{lst:dependency}, figure \ref{fig:dependency} shows the corresponding dependency graph that results, where every node is a statement of one of the three threads, and the arrows the dependency of that statement (e.g., $\alpha1$ must happen before $\alpha2$). The extra challenge is that this dependency graph is effectively post-mortem, but the runtime system needs to be able to build and solve these graphs as the dependencies unfold. Resolving dependency graphs being a complex and expensive endeavour, this solution is not the preferred one.
1298
1299\subsubsection{Partial Signalling} \label{partial-sig}
1300Finally, the solution that is chosen for \CFA is to use partial signalling. Again using listing \ref{lst:int-bulk-pseudo}, the partial signalling solution transfers ownership of monitor \code{B} at lines \ref{line:signal1} to the waiter but does not wake the waiting thread since it is still using monitor \code{A}. Only when it reaches line \ref{line:lastRelease} does it actually wake up the waiting thread. This solution has the benefit that complexity is encapsulated into only two actions: passing monitors to the next owner when they should be released and conditionally waking threads if all conditions are met. This solution has a much simpler implementation than a dependency graph solving algorithms, which is why it was chosen. Furthermore, after being fully implemented, this solution does not appear to have any significant downsides.
1301
1302Using partial signalling, listing \ref{lst:dependency} can be solved easily:
1303\begin{itemize}
1304        \item When thread $\gamma$ reaches line \ref{line:release-ab} it transfers monitor \code{B} to thread $\alpha$ and continues to hold monitor \code{A}.
1305        \item When thread $\gamma$ reaches line \ref{line:release-a}  it transfers monitor \code{A} to thread $\beta$  and wakes it up.
1306        \item When thread $\beta$  reaches line \ref{line:release-aa} it transfers monitor \code{A} to thread $\alpha$ and wakes it up.
1307\end{itemize}
1308
1309% ======================================================================
1310% ======================================================================
1311\subsection{Signalling: Now or Later}
1312% ======================================================================
1313% ======================================================================
1314\begin{table}
1315\begin{tabular}{|c|c|}
1316\code{signal} & \code{signal_block} \\
1317\hline
1318\begin{cfacode}[tabsize=3]
1319monitor DatingService
1320{
1321        //compatibility codes
1322        enum{ CCodes = 20 };
1323
1324        int girlPhoneNo
1325        int boyPhoneNo;
1326};
1327
1328condition girls[CCodes];
1329condition boys [CCodes];
1330condition exchange;
1331
1332int girl(int phoneNo, int ccode)
1333{
1334        //no compatible boy ?
1335        if(empty(boys[ccode]))
1336        {
1337                //wait for boy
1338                wait(girls[ccode]);
1339
1340                //make phone number available
1341                girlPhoneNo = phoneNo;
1342
1343                //wake boy from chair
1344                signal(exchange);
1345        }
1346        else
1347        {
1348                //make phone number available
1349                girlPhoneNo = phoneNo;
1350
1351                //wake boy
1352                signal(boys[ccode]);
1353
1354                //sit in chair
1355                wait(exchange);
1356        }
1357        return boyPhoneNo;
1358}
1359
1360int boy(int phoneNo, int ccode)
1361{
1362        //same as above
1363        //with boy/girl interchanged
1364}
1365\end{cfacode}&\begin{cfacode}[tabsize=3]
1366monitor DatingService
1367{
1368        //compatibility codes
1369        enum{ CCodes = 20 };
1370
1371        int girlPhoneNo;
1372        int boyPhoneNo;
1373};
1374
1375condition girls[CCodes];
1376condition boys [CCodes];
1377//exchange is not needed
1378
1379int girl(int phoneNo, int ccode)
1380{
1381        //no compatible boy ?
1382        if(empty(boys[ccode]))
1383        {
1384                //wait for boy
1385                wait(girls[ccode]);
1386
1387                //make phone number available
1388                girlPhoneNo = phoneNo;
1389
1390                //wake boy from chair
1391                signal(exchange);
1392        }
1393        else
1394        {
1395                //make phone number available
1396                girlPhoneNo = phoneNo;
1397
1398                //wake boy
1399                signal_block(boys[ccode]);
1400
1401                //second handshake unnecessary
1402
1403        }
1404        return boyPhoneNo;
1405}
1406
1407int boy(int phoneNo, int ccode)
1408{
1409        //same as above
1410        //with boy/girl interchanged
1411}
1412\end{cfacode}
1413\end{tabular}
1414\caption{Dating service example using \code{signal} and \code{signal_block}. }
1415\label{tbl:datingservice}
1416\end{table}
1417An important note is that, until now, signalling a monitor was a delayed operation. The ownership of the monitor is transferred only when the monitor would have otherwise been released, not at the point of the \code{signal} statement. However, in some cases, it may be more convenient for users to immediately transfer ownership to the thread that is waiting for cooperation, which is achieved using the \code{signal_block} routine.
1418
1419The example in table \ref{tbl:datingservice} highlights the difference in behaviour. As mentioned, \code{signal} only transfers ownership once the current critical section exits; this behaviour requires additional synchronization when a two-way handshake is needed. To avoid this explicit synchronization, the \code{condition} type offers the \code{signal_block} routine, which handles the two-way handshake as shown in the example. This feature removes the need for a second condition variables and simplifies programming. Like every other monitor semantic, \code{signal_block} uses barging prevention, which means mutual-exclusion is baton-passed both on the front end and the back end of the call to \code{signal_block}, meaning no other thread can acquire the monitor either before or after the call.
1420
1421% ======================================================================
1422% ======================================================================
1423\section{External scheduling} \label{extsched}
1424% ======================================================================
1425% ======================================================================
1426An alternative to internal scheduling is external scheduling (see Table~\ref{tbl:sched}).
1427\begin{table}
1428\begin{tabular}{|c|c|c|}
1429Internal Scheduling & External Scheduling & Go\\
1430\hline
1431\begin{ucppcode}[tabsize=3]
1432_Monitor Semaphore {
1433        condition c;
1434        bool inUse;
1435public:
1436        void P() {
1437                if(inUse)
1438                        wait(c);
1439                inUse = true;
1440        }
1441        void V() {
1442                inUse = false;
1443                signal(c);
1444        }
1445}
1446\end{ucppcode}&\begin{ucppcode}[tabsize=3]
1447_Monitor Semaphore {
1448
1449        bool inUse;
1450public:
1451        void P() {
1452                if(inUse)
1453                        _Accept(V);
1454                inUse = true;
1455        }
1456        void V() {
1457                inUse = false;
1458
1459        }
1460}
1461\end{ucppcode}&\begin{gocode}[tabsize=3]
1462type MySem struct {
1463        inUse bool
1464        c     chan bool
1465}
1466
1467// acquire
1468func (s MySem) P() {
1469        if s.inUse {
1470                select {
1471                case <-s.c:
1472                }
1473        }
1474        s.inUse = true
1475}
1476
1477// release
1478func (s MySem) V() {
1479        s.inUse = false
1480
1483        s.c <- false
1484}
1485\end{gocode}
1486\end{tabular}
1487\caption{Different forms of scheduling.}
1488\label{tbl:sched}
1489\end{table}
1490This method is more constrained and explicit, which helps users reduce the non-deterministic nature of concurrency. Indeed, as the following examples demonstrate, external scheduling allows users to wait for events from other threads without the concern of unrelated events occurring. External scheduling can generally be done either in terms of control flow (e.g., Ada with \code{accept}, \uC with \code{_Accept}) or in terms of data (e.g., Go with channels). Of course, both of these paradigms have their own strengths and weaknesses, but for this project, control-flow semantics was chosen to stay consistent with the rest of the languages semantics. Two challenges specific to \CFA arise when trying to add external scheduling with loose object definitions and multiple-monitor routines. The previous example shows a simple use \code{_Accept} versus \code{wait}/\code{signal} and its advantages. Note that while other languages often use \code{accept}/\code{select} as the core external scheduling keyword, \CFA uses \code{waitfor} to prevent name collisions with existing socket \textbf{api}s.
1491
1492For the \code{P} member above using internal scheduling, the call to \code{wait} only guarantees that \code{V} is the last routine to access the monitor, allowing a third routine, say \code{isInUse()}, acquire mutual exclusion several times while routine \code{P} is waiting. On the other hand, external scheduling guarantees that while routine \code{P} is waiting, no other routine than \code{V} can acquire the monitor.
1493
1494% ======================================================================
1495% ======================================================================
1496\subsection{Loose Object Definitions}
1497% ======================================================================
1498% ======================================================================
1499In \uC, a monitor class declaration includes an exhaustive list of monitor operations. Since \CFA is not object oriented, monitors become both more difficult to implement and less clear for a user:
1500
1501\begin{cfacode}
1502monitor A {};
1503
1504void f(A & mutex a);
1505void g(A & mutex a) {
1506        waitfor(f); //Obvious which f() to wait for
1507}
1508
1509void f(A & mutex a, int); //New different F added in scope
1510void h(A & mutex a) {
1511        waitfor(f); //Less obvious which f() to wait for
1512}
1513\end{cfacode}
1514
1515Furthermore, external scheduling is an example where implementation constraints become visible from the interface. Here is the pseudo-code for the entering phase of a monitor:
1516\begin{center}
1517\begin{tabular}{l}
1518\begin{pseudo}
1519        if monitor is free
1520                enter
1521        elif already own the monitor
1522                continue
1523        elif monitor accepts me
1524                enter
1525        else
1526                block
1527\end{pseudo}
1528\end{tabular}
1529\end{center}
1530For the first two conditions, it is easy to implement a check that can evaluate the condition in a few instructions. However, a fast check for \pscode{monitor accepts me} is much harder to implement depending on the constraints put on the monitors. Indeed, monitors are often expressed as an entry queue and some acceptor queue as in Figure~\ref{fig:ClassicalMonitor}.
1531
1532\begin{figure}
1533\centering
1534\subfloat[Classical Monitor] {
1535\label{fig:ClassicalMonitor}
1536{\resizebox{0.45\textwidth}{!}{\input{monitor}}}
1537}% subfloat
1539\subfloat[\textbf{bulk-acq} Monitor] {
1540\label{fig:BulkMonitor}
1541{\resizebox{0.45\textwidth}{!}{\input{ext_monitor}}}
1542}% subfloat
1543\caption{External Scheduling Monitor}
1544\end{figure}
1545
1546There are other alternatives to these pictures, but in the case of the left picture, implementing a fast accept check is relatively easy. Restricted to a fixed number of mutex members, N, the accept check reduces to updating a bitmask when the acceptor queue changes, a check that executes in a single instruction even with a fairly large number (e.g., 128) of mutex members. This approach requires a unique dense ordering of routines with an upper-bound and that ordering must be consistent across translation units. For OO languages these constraints are common, since objects only offer adding member routines consistently across translation units via inheritance. However, in \CFA users can extend objects with mutex routines that are only visible in certain translation unit. This means that establishing a program-wide dense-ordering among mutex routines can only be done in the program linking phase, and still could have issues when using dynamically shared objects.
1547
1548The alternative is to alter the implementation as in Figure~\ref{fig:BulkMonitor}.
1549Here, the mutex routine called is associated with a thread on the entry queue while a list of acceptable routines is kept separate. Generating a mask dynamically means that the storage for the mask information can vary between calls to \code{waitfor}, allowing for more flexibility and extensions. Storing an array of accepted function pointers replaces the single instruction bitmask comparison with dereferencing a pointer followed by a linear search. Furthermore, supporting nested external scheduling (e.g., listing \ref{lst:nest-ext}) may now require additional searches for the \code{waitfor} statement to check if a routine is already queued.
1550
1551\begin{figure}
1552\begin{cfacode}[caption={Example of nested external scheduling},label={lst:nest-ext}]
1553monitor M {};
1554void foo( M & mutex a ) {}
1555void bar( M & mutex b ) {
1556        //Nested in the waitfor(bar, c) call
1557        waitfor(foo, b);
1558}
1559void baz( M & mutex c ) {
1560        waitfor(bar, c);
1561}
1562
1563\end{cfacode}
1564\end{figure}
1565
1566Note that in the right picture, tasks need to always keep track of the monitors associated with mutex routines, and the routine mask needs to have both a function pointer and a set of monitors, as is discussed in the next section. These details are omitted from the picture for the sake of simplicity.
1567
1568At this point, a decision must be made between flexibility and performance. Many design decisions in \CFA achieve both flexibility and performance, for example polymorphic routines add significant flexibility but inlining them means the optimizer can easily remove any runtime cost. Here, however, the cost of flexibility cannot be trivially removed. In the end, the most flexible approach has been chosen since it allows users to write programs that would otherwise be  hard to write. This decision is based on the assumption that writing fast but inflexible locks is closer to a solved problem than writing locks that are as flexible as external scheduling in \CFA.
1569
1570% ======================================================================
1571% ======================================================================
1572\subsection{Multi-Monitor Scheduling}
1573% ======================================================================
1574% ======================================================================
1575
1576External scheduling, like internal scheduling, becomes significantly more complex when introducing multi-monitor syntax. Even in the simplest possible case, some new semantics needs to be established:
1577\begin{cfacode}
1578monitor M {};
1579
1580void f(M & mutex a);
1581
1582void g(M & mutex b, M & mutex c) {
1583        waitfor(f); //two monitors M => unknown which to pass to f(M & mutex)
1584}
1585\end{cfacode}
1586The obvious solution is to specify the correct monitor as follows:
1587
1588\begin{cfacode}
1589monitor M {};
1590
1591void f(M & mutex a);
1592
1593void g(M & mutex a, M & mutex b) {
1594        //wait for call to f with argument b
1595        waitfor(f, b);
1596}
1597\end{cfacode}
1598This syntax is unambiguous. Both locks are acquired and kept by \code{g}. When routine \code{f} is called, the lock for monitor \code{b} is temporarily transferred from \code{g} to \code{f} (while \code{g} still holds lock \code{a}). This behaviour can be extended to the multi-monitor \code{waitfor} statement as follows.
1599
1600\begin{cfacode}
1601monitor M {};
1602
1603void f(M & mutex a, M & mutex b);
1604
1605void g(M & mutex a, M & mutex b) {
1606        //wait for call to f with arguments a and b
1607        waitfor(f, a, b);
1608}
1609\end{cfacode}
1610
1611Note that the set of monitors passed to the \code{waitfor} statement must be entirely contained in the set of monitors already acquired in the routine. \code{waitfor} used in any other context is undefined behaviour.
1612
1613An important behaviour to note is when a set of monitors only match partially:
1614
1615\begin{cfacode}
1616mutex struct A {};
1617
1618mutex struct B {};
1619
1620void g(A & mutex a, B & mutex b) {
1621        waitfor(f, a, b);
1622}
1623
1624A a1, a2;
1625B b;
1626
1627void foo() {
1628        g(a1, b); //block on accept
1629}
1630
1631void bar() {
1632        f(a2, b); //fulfill cooperation
1633}
1634\end{cfacode}
1635While the equivalent can happen when using internal scheduling, the fact that conditions are specific to a set of monitors means that users have to use two different condition variables. In both cases, partially matching monitor sets does not wakeup the waiting thread. It is also important to note that in the case of external scheduling the order of parameters is irrelevant; \code{waitfor(f,a,b)} and \code{waitfor(f,b,a)} are indistinguishable waiting condition.
1636
1637% ======================================================================
1638% ======================================================================
1639\subsection{\code{waitfor} Semantics}
1640% ======================================================================
1641% ======================================================================
1642
1643Syntactically, the \code{waitfor} statement takes a function identifier and a set of monitors. While the set of monitors can be any list of expressions, the function name is more restricted because the compiler validates at compile time the validity of the function type and the parameters used with the \code{waitfor} statement. It checks that the set of monitors passed in matches the requirements for a function call. Listing \ref{lst:waitfor} shows various usages of the waitfor statement and which are acceptable. The choice of the function type is made ignoring any non-\code{mutex} parameter. One limitation of the current implementation is that it does not handle overloading, but overloading is possible.
1644\begin{figure}
1645\begin{cfacode}[caption={Various correct and incorrect uses of the waitfor statement},label={lst:waitfor}]
1646monitor A{};
1647monitor B{};
1648
1649void f1( A & mutex );
1650void f2( A & mutex, B & mutex );
1651void f3( A & mutex, int );
1652void f4( A & mutex, int );
1653void f4( A & mutex, double );
1654
1655void foo( A & mutex a1, A & mutex a2, B & mutex b1, B & b2 ) {
1656        A * ap = & a1;
1657        void (*fp)( A & mutex ) = f1;
1658
1659        waitfor(f1, a1);     //Correct : 1 monitor case
1660        waitfor(f2, a1, b1); //Correct : 2 monitor case
1661        waitfor(f3, a1);     //Correct : non-mutex arguments are ignored
1662        waitfor(f1, *ap);    //Correct : expression as argument
1663
1664        waitfor(f1, a1, b1); //Incorrect : Too many mutex arguments
1665        waitfor(f2, a1);     //Incorrect : Too few mutex arguments
1666        waitfor(f2, a1, a2); //Incorrect : Mutex arguments don't match
1667        waitfor(f1, 1);      //Incorrect : 1 not a mutex argument
1668        waitfor(f9, a1);     //Incorrect : f9 function does not exist
1669        waitfor(*fp, a1 );   //Incorrect : fp not an identifier
1670        waitfor(f4, a1);     //Incorrect : f4 ambiguous
1671
1672        waitfor(f2, a1, b2); //Undefined behaviour : b2 not mutex
1673}
1674\end{cfacode}
1675\end{figure}
1676
1677Finally, for added flexibility, \CFA supports constructing a complex \code{waitfor} statement using the \code{or}, \code{timeout} and \code{else}. Indeed, multiple \code{waitfor} clauses can be chained together using \code{or}; this chain forms a single statement that uses baton pass to any function that fits one of the function+monitor set passed in. To enable users to tell which accepted function executed, \code{waitfor}s are followed by a statement (including the null statement \code{;}) or a compound statement, which is executed after the clause is triggered. A \code{waitfor} chain can also be followed by a \code{timeout}, to signify an upper bound on the wait, or an \code{else}, to signify that the call should be non-blocking, which checks for a matching function call already arrived and otherwise continues. Any and all of these clauses can be preceded by a \code{when} condition to dynamically toggle the accept clauses on or off based on some current state. Listing \ref{lst:waitfor2} demonstrates several complex masks and some incorrect ones.
1678
1679\begin{figure}
1680\begin{cfacode}[caption={Various correct and incorrect uses of the or, else, and timeout clause around a waitfor statement},label={lst:waitfor2}]
1681monitor A{};
1682
1683void f1( A & mutex );
1684void f2( A & mutex );
1685
1686void foo( A & mutex a, bool b, int t ) {
1687        //Correct : blocking case
1688        waitfor(f1, a);
1689
1690        //Correct : block with statement
1691        waitfor(f1, a) {
1692                sout | "f1" | endl;
1693        }
1694
1695        //Correct : block waiting for f1 or f2
1696        waitfor(f1, a) {
1697                sout | "f1" | endl;
1698        } or waitfor(f2, a) {
1699                sout | "f2" | endl;
1700        }
1701
1702        //Correct : non-blocking case
1703        waitfor(f1, a); or else;
1704
1705        //Correct : non-blocking case
1706        waitfor(f1, a) {
1707                sout | "blocked" | endl;
1708        } or else {
1709                sout | "didn't block" | endl;
1710        }
1711
1712        //Correct : block at most 10 seconds
1713        waitfor(f1, a) {
1714                sout | "blocked" | endl;
1715        } or timeout( 10s) {
1716                sout | "didn't block" | endl;
1717        }
1718
1719        //Correct : block only if b == true
1720        //if b == false, don't even make the call
1721        when(b) waitfor(f1, a);
1722
1723        //Correct : block only if b == true
1724        //if b == false, make non-blocking call
1725        waitfor(f1, a); or when(!b) else;
1726
1727        //Correct : block only of t > 1
1728        waitfor(f1, a); or when(t > 1) timeout(t); or else;
1729
1730        //Incorrect : timeout clause is dead code
1731        waitfor(f1, a); or timeout(t); or else;
1732
1733        //Incorrect : order must be
1734        //waitfor [or waitfor... [or timeout] [or else]]
1735        timeout(t); or waitfor(f1, a); or else;
1736}
1737\end{cfacode}
1738\end{figure}
1739
1740% ======================================================================
1741% ======================================================================
1742\subsection{Waiting For The Destructor}
1743% ======================================================================
1744% ======================================================================
1745An interesting use for the \code{waitfor} statement is destructor semantics. Indeed, the \code{waitfor} statement can accept any \code{mutex} routine, which includes the destructor (see section \ref{data}). However, with the semantics discussed until now, waiting for the destructor does not make any sense, since using an object after its destructor is called is undefined behaviour. The simplest approach is to disallow \code{waitfor} on a destructor. However, a more expressive approach is to flip ordering of execution when waiting for the destructor, meaning that waiting for the destructor allows the destructor to run after the current \code{mutex} routine, similarly to how a condition is signalled.
1746\begin{figure}
1747\begin{cfacode}[caption={Example of an executor which executes action in series until the destructor is called.},label={lst:dtor-order}]
1748monitor Executer {};
1749struct  Action;
1750
1751void ^?{}   (Executer & mutex this);
1752void execute(Executer & mutex this, const Action & );
1753void run    (Executer & mutex this) {
1754        while(true) {
1755                   waitfor(execute, this);
1756                or waitfor(^?{}   , this) {
1757                        break;
1758                }
1759        }
1760}
1761\end{cfacode}
1762\end{figure}
1763For example, listing \ref{lst:dtor-order} shows an example of an executor with an infinite loop, which waits for the destructor to break out of this loop. Switching the semantic meaning introduces an idiomatic way to terminate a task and/or wait for its termination via destruction.
1764
1765
1766% ######     #    ######     #    #       #       ####### #       ###  #####  #     #
1767% #     #   # #   #     #   # #   #       #       #       #        #  #     # ##   ##
1768% #     #  #   #  #     #  #   #  #       #       #       #        #  #       # # # #
1769% ######  #     # ######  #     # #       #       #####   #        #   #####  #  #  #
1770% #       ####### #   #   ####### #       #       #       #        #        # #     #
1771% #       #     # #    #  #     # #       #       #       #        #  #     # #     #
1772% #       #     # #     # #     # ####### ####### ####### ####### ###  #####  #     #
1773\section{Parallelism}
1774Historically, computer performance was about processor speeds and instruction counts. However, with heat dissipation being a direct consequence of speed increase, parallelism has become the new source for increased performance~\cite{Sutter05, Sutter05b}. In this decade, it is no longer reasonable to create a high-performance application without caring about parallelism. Indeed, parallelism is an important aspect of performance and more specifically throughput and hardware utilization. The lowest-level approach of parallelism is to use \textbf{kthread} in combination with semantics like \code{fork}, \code{join}, etc. However, since these have significant costs and limitations, \textbf{kthread} are now mostly used as an implementation tool rather than a user oriented one. There are several alternatives to solve these issues that all have strengths and weaknesses. While there are many variations of the presented paradigms, most of these variations do not actually change the guarantees or the semantics, they simply move costs in order to achieve better performance for certain workloads.
1775
1778A direct improvement on the \textbf{kthread} approach is to use \textbf{uthread}. These threads offer most of the same features that the operating system already provides but can be used on a much larger scale. This approach is the most powerful solution as it allows all the features of multithreading, while removing several of the more expensive costs of kernel threads. The downside is that almost none of the low-level threading problems are hidden; users still have to think about data races, deadlocks and synchronization issues. These issues can be somewhat alleviated by a concurrency toolkit with strong guarantees, but the parallelism toolkit offers very little to reduce complexity in itself.
1779
1780Examples of languages that support \textbf{uthread} are Erlang~\cite{Erlang} and \uC~\cite{uC++book}.
1781
1782\subsection{Fibers : User-Level Threads Without Preemption} \label{fibers}
1783A popular variant of \textbf{uthread} is what is often referred to as \textbf{fiber}. However, \textbf{fiber} do not present meaningful semantic differences with \textbf{uthread}. The significant difference between \textbf{uthread} and \textbf{fiber} is the lack of \textbf{preemption} in the latter. Advocates of \textbf{fiber} list their high performance and ease of implementation as major strengths, but the performance difference between \textbf{uthread} and \textbf{fiber} is controversial, and the ease of implementation, while true, is a weak argument in the context of language design. Therefore this proposal largely ignores fibers.
1784
1785An example of a language that uses fibers is Go~\cite{Go}
1786
1788An approach on the opposite end of the spectrum is to base parallelism on \textbf{pool}. Indeed, \textbf{pool} offer limited flexibility but at the benefit of a simpler user interface. In \textbf{pool} based systems, users express parallelism as units of work, called jobs, and a dependency graph (either explicit or implicit) that ties them together. This approach means users need not worry about concurrency but significantly limit the interaction that can occur among jobs. Indeed, any \textbf{job} that blocks also block the underlying worker, which effectively means the CPU utilization, and therefore throughput, suffers noticeably. It can be argued that a solution to this problem is to use more workers than available cores. However, unless the number of jobs and the number of workers are comparable, having a significant number of blocked jobs always results in idles cores.
1789
1790The gold standard of this implementation is Intel's TBB library~\cite{TBB}.
1791
1793While the choice between the three paradigms listed above may have significant performance implications, it is difficult to pin down the performance implications of choosing a model at the language level. Indeed, in many situations one of these paradigms may show better performance but it all strongly depends on the workload. Having a large amount of mostly independent units of work to execute almost guarantees equivalent performance across paradigms and that the \textbf{pool}-based system has the best efficiency thanks to the lower memory overhead (i.e., no thread stack per job). However, interactions among jobs can easily exacerbate contention. User-level threads allow fine-grain context switching, which results in better resource utilization, but a context switch is more expensive and the extra control means users need to tweak more variables to get the desired performance. Finally, if the units of uninterrupted work are large, enough the paradigm choice is largely amortized by the actual work done.
1794
1795\section{The \protect\CFA\ Kernel : Processors, Clusters and Threads}\label{kernel}
1796A \textbf{cfacluster} is a group of \textbf{kthread} executed in isolation. \textbf{uthread} are scheduled on the \textbf{kthread} of a given \textbf{cfacluster}, allowing organization between \textbf{uthread} and \textbf{kthread}. It is important that \textbf{kthread} belonging to a same \textbf{cfacluster} have homogeneous settings, otherwise migrating a \textbf{uthread} from one \textbf{kthread} to the other can cause issues. A \textbf{cfacluster} also offers a pluggable scheduler that can optimize the workload generated by the \textbf{uthread}.
1797
1798\textbf{cfacluster} have not been fully implemented in the context of this paper. Currently \CFA only supports one \textbf{cfacluster}, the initial one.
1799
1800\subsection{Future Work: Machine Setup}\label{machine}
1801While this was not done in the context of this paper, another important aspect of clusters is affinity. While many common desktop and laptop PCs have homogeneous CPUs, other devices often have more heterogeneous setups. For example, a system using \textbf{numa} configurations may benefit from users being able to tie clusters and/or kernel threads to certain CPU cores. OS support for CPU affinity is now common~\cite{affinityLinux, affinityWindows, affinityFreebsd, affinityNetbsd, affinityMacosx}, which means it is both possible and desirable for \CFA to offer an abstraction mechanism for portable CPU affinity.
1802
1804Given these building blocks, it is possible to reproduce all three of the popular paradigms. Indeed, \textbf{uthread} is the default paradigm in \CFA. However, disabling \textbf{preemption} on the \textbf{cfacluster} means \textbf{cfathread} effectively become \textbf{fiber}. Since several \textbf{cfacluster} with different scheduling policy can coexist in the same application, this allows \textbf{fiber} and \textbf{uthread} to coexist in the runtime of an application. Finally, it is possible to build executors for thread pools from \textbf{uthread} or \textbf{fiber}, which includes specialized jobs like actors~\cite{Actors}.
1805
1806
1807
1808\section{Behind the Scenes}
1809There are several challenges specific to \CFA when implementing concurrency. These challenges are a direct result of \textbf{bulk-acq} and loose object definitions. These two constraints are the root cause of most design decisions in the implementation. Furthermore, to avoid contention from dynamically allocating memory in a concurrent environment, the internal-scheduling design is (almost) entirely free of mallocs. This approach avoids the chicken and egg problem~\cite{Chicken} of having a memory allocator that relies on the threading system and a threading system that relies on the runtime. This extra goal means that memory management is a constant concern in the design of the system.
1810
1811The main memory concern for concurrency is queues. All blocking operations are made by parking threads onto queues and all queues are designed with intrusive nodes, where each node has pre-allocated link fields for chaining, to avoid the need for memory allocation. Since several concurrency operations can use an unbound amount of memory (depending on \textbf{bulk-acq}), statically defining information in the intrusive fields of threads is insufficient.The only way to use a variable amount of memory without requiring memory allocation is to pre-allocate large buffers of memory eagerly and store the information in these buffers. Conveniently, the call stack fits that description and is easy to use, which is why it is used heavily in the implementation of internal scheduling, particularly variable-length arrays. Since stack allocation is based on scopes, the first step of the implementation is to identify the scopes that are available to store the information, and which of these can have a variable-length array. The threads and the condition both have a fixed amount of memory, while \code{mutex} routines and blocking calls allow for an unbound amount, within the stack size.
1812
1813Note that since the major contributions of this paper are extending monitor semantics to \textbf{bulk-acq} and loose object definitions, any challenges that are not resulting of these characteristics of \CFA are considered as solved problems and therefore not discussed.
1814
1815% ======================================================================
1816% ======================================================================
1817\section{Mutex Routines}
1818% ======================================================================
1819% ======================================================================
1820
1821The first step towards the monitor implementation is simple \code{mutex} routines. In the single monitor case, mutual-exclusion is done using the entry/exit procedure in listing \ref{lst:entry1}. The entry/exit procedures do not have to be extended to support multiple monitors. Indeed it is sufficient to enter/leave monitors one-by-one as long as the order is correct to prevent deadlock~\cite{Havender68}. In \CFA, ordering of monitor acquisition relies on memory ordering. This approach is sufficient because all objects are guaranteed to have distinct non-overlapping memory layouts and mutual-exclusion for a monitor is only defined for its lifetime, meaning that destroying a monitor while it is acquired is undefined behaviour. When a mutex call is made, the concerned monitors are aggregated into a variable-length pointer array and sorted based on pointer values. This array persists for the entire duration of the mutual-exclusion and its ordering reused extensively.
1822\begin{figure}
1823\begin{multicols}{2}
1824Entry
1825\begin{pseudo}
1826if monitor is free
1827        enter
1829        continue
1830else
1831        block
1832increment recursions
1833\end{pseudo}
1834\columnbreak
1835Exit
1836\begin{pseudo}
1837decrement recursion
1838if recursion == 0
1839        if entry queue not empty
1841\end{pseudo}
1842\end{multicols}
1843\begin{pseudo}[caption={Initial entry and exit routine for monitors},label={lst:entry1}]
1844\end{pseudo}
1845\end{figure}
1846
1847\subsection{Details: Interaction with polymorphism}
1848Depending on the choice of semantics for when monitor locks are acquired, interaction between monitors and \CFA's concept of polymorphism can be more complex to support. However, it is shown that entry-point locking solves most of the issues.
1849
1850First of all, interaction between \code{otype} polymorphism (see Section~\ref{s:ParametricPolymorphism}) and monitors is impossible since monitors do not support copying. Therefore, the main question is how to support \code{dtype} polymorphism. It is important to present the difference between the two acquiring options: \textbf{callsite-locking} and entry-point locking, i.e., acquiring the monitors before making a mutex routine-call or as the first operation of the mutex routine-call. For example:
1851\begin{table}[H]
1852\begin{center}
1853\begin{tabular}{|c|c|c|}
1854Mutex & \textbf{callsite-locking} & \textbf{entry-point-locking} \\
1855call & pseudo-code & pseudo-code \\
1856\hline
1857\begin{cfacode}[tabsize=3]
1858void foo(monitor& mutex a){
1859
1860        //Do Work
1861        //...
1862
1863}
1864
1865void main() {
1866        monitor a;
1867
1868        foo(a);
1869
1870}
1871\end{cfacode} & \begin{pseudo}[tabsize=3]
1872foo(& a) {
1873
1874        //Do Work
1875        //...
1876
1877}
1878
1879main() {
1880        monitor a;
1881        acquire(a);
1882        foo(a);
1883        release(a);
1884}
1885\end{pseudo} & \begin{pseudo}[tabsize=3]
1886foo(& a) {
1887        acquire(a);
1888        //Do Work
1889        //...
1890        release(a);
1891}
1892
1893main() {
1894        monitor a;
1895
1896        foo(a);
1897
1898}
1899\end{pseudo}
1900\end{tabular}
1901\end{center}
1902\caption{Call-site vs entry-point locking for mutex calls}
1903\label{tbl:locking-site}
1904\end{table}
1905
1906Note the \code{mutex} keyword relies on the type system, which means that in cases where a generic monitor-routine is desired, writing the mutex routine is possible with the proper trait, e.g.:
1907\begin{cfacode}
1908//Incorrect: T may not be monitor
1909forall(dtype T)
1910void foo(T * mutex t);
1911
1912//Correct: this function only works on monitors (any monitor)
1913forall(dtype T | is_monitor(T))
1914void bar(T * mutex t));
1915\end{cfacode}
1916
1917Both entry point and \textbf{callsite-locking} are feasible implementations. The current \CFA implementation uses entry-point locking because it requires less work when using \textbf{raii}, effectively transferring the burden of implementation to object construction/destruction. It is harder to use \textbf{raii} for call-site locking, as it does not necessarily have an existing scope that matches exactly the scope of the mutual exclusion, i.e., the function body. For example, the monitor call can appear in the middle of an expression. Furthermore, entry-point locking requires less code generation since any useful routine is called multiple times but there is only one entry point for many call sites.
1918
1919% ======================================================================
1920% ======================================================================
1922% ======================================================================
1923% ======================================================================
1924
1925Figure \ref{fig:system1} shows a high-level picture if the \CFA runtime system in regards to concurrency. Each component of the picture is explained in detail in the flowing sections.
1926
1927\begin{figure}
1928\begin{center}
1929{\resizebox{\textwidth}{!}{\input{system.pstex_t}}}
1930\end{center}
1931\caption{Overview of the entire system}
1932\label{fig:system1}
1933\end{figure}
1934
1935\subsection{Processors}
1936Parallelism in \CFA is built around using processors to specify how much parallelism is desired. \CFA processors are object wrappers around kernel threads, specifically \texttt{pthread}s in the current implementation of \CFA. Indeed, any parallelism must go through operating-system libraries. However, \textbf{uthread} are still the main source of concurrency, processors are simply the underlying source of parallelism. Indeed, processor \textbf{kthread} simply fetch a \textbf{uthread} from the scheduler and run it; they are effectively executers for user-threads. The main benefit of this approach is that it offers a well-defined boundary between kernel code and user code, for example, kernel thread quiescing, scheduling and interrupt handling. Processors internally use coroutines to take advantage of the existing context-switching semantics.
1937
1938\subsection{Stack Management}
1939One of the challenges of this system is to reduce the footprint as much as possible. Specifically, all \texttt{pthread}s created also have a stack created with them, which should be used as much as possible. Normally, coroutines also create their own stack to run on, however, in the case of the coroutines used for processors, these coroutines run directly on the \textbf{kthread} stack, effectively stealing the processor stack. The exception to this rule is the Main Processor, i.e., the initial \textbf{kthread} that is given to any program. In order to respect C user expectations, the stack of the initial kernel thread, the main stack of the program, is used by the main user thread rather than the main processor, which can grow very large.
1940
1941\subsection{Context Switching}
1942As mentioned in section \ref{coroutine}, coroutines are a stepping stone for implementing threading, because they share the same mechanism for context-switching between different stacks. To improve performance and simplicity, context-switching is implemented using the following assumption: all context-switches happen inside a specific function call. This assumption means that the context-switch only has to copy the callee-saved registers onto the stack and then switch the stack registers with the ones of the target coroutine/thread. Note that the instruction pointer can be left untouched since the context-switch is always inside the same function. Threads, however, do not context-switch between each other directly. They context-switch to the scheduler. This method is called a 2-step context-switch and has the advantage of having a clear distinction between user code and the kernel where scheduling and other system operations happen. Obviously, this doubles the context-switch cost because threads must context-switch to an intermediate stack. The alternative 1-step context-switch uses the stack of the from'' thread to schedule and then context-switches directly to the to'' thread. However, the performance of the 2-step context-switch is still superior to a \code{pthread_yield} (see section \ref{results}). Additionally, for users in need for optimal performance, it is important to note that having a 2-step context-switch as the default does not prevent \CFA from offering a 1-step context-switch (akin to the Microsoft \code{SwitchToFiber}~\cite{switchToWindows} routine). This option is not currently present in \CFA, but the changes required to add it are strictly additive.
1943
1944\subsection{Preemption} \label{preemption}
1945Finally, an important aspect for any complete threading system is preemption. As mentioned in section \ref{basics}, preemption introduces an extra degree of uncertainty, which enables users to have multiple threads interleave transparently, rather than having to cooperate among threads for proper scheduling and CPU distribution. Indeed, preemption is desirable because it adds a degree of isolation among threads. In a fully cooperative system, any thread that runs a long loop can starve other threads, while in a preemptive system, starvation can still occur but it does not rely on every thread having to yield or block on a regular basis, which reduces significantly a programmer burden. Obviously, preemption is not optimal for every workload. However any preemptive system can become a cooperative system by making the time slices extremely large. Therefore, \CFA uses a preemptive threading system.
1946
1947Preemption in \CFA\footnote{Note that the implementation of preemption is strongly tied with the underlying threading system. For this reason, only the Linux implementation is cover, \CFA does not run on Windows at the time of writting} is based on kernel timers, which are used to run a discrete-event simulation. Every processor keeps track of the current time and registers an expiration time with the preemption system. When the preemption system receives a change in preemption, it inserts the time in a sorted order and sets a kernel timer for the closest one, effectively stepping through preemption events on each signal sent by the timer. These timers use the Linux signal {\tt SIGALRM}, which is delivered to the process rather than the kernel-thread. This results in an implementation problem, because when delivering signals to a process, the kernel can deliver the signal to any kernel thread for which the signal is not blocked, i.e.:
1948\begin{quote}
1949A process-directed signal may be delivered to any one of the threads that does not currently have the signal blocked. If more than one of the threads has the signal unblocked, then the kernel chooses an arbitrary thread to which to deliver the signal.
1950SIGNAL(7) - Linux Programmer's Manual
1951\end{quote}
1952For the sake of simplicity, and in order to prevent the case of having two threads receiving alarms simultaneously, \CFA programs block the {\tt SIGALRM} signal on every kernel thread except one.
1953
1954Now because of how involuntary context-switches are handled, the kernel thread handling {\tt SIGALRM} cannot also be a processor thread. Hence, involuntary context-switching is done by sending signal {\tt SIGUSR1} to the corresponding proces\-sor and having the thread yield from inside the signal handler. This approach effectively context-switches away from the signal handler back to the kernel and the signal handler frame is eventually unwound when the thread is scheduled again. As a result, a signal handler can start on one kernel thread and terminate on a second kernel thread (but the same user thread). It is important to note that signal handlers save and restore signal masks because user-thread migration can cause a signal mask to migrate from one kernel thread to another. This behaviour is only a problem if all kernel threads, among which a user thread can migrate, differ in terms of signal masks\footnote{Sadly, official POSIX documentation is silent on what distinguishes async-signal-safe'' functions from other functions.}. However, since the kernel thread handling preemption requires a different signal mask, executing user threads on the kernel-alarm thread can cause deadlocks. For this reason, the alarm thread is in a tight loop around a system call to \code{sigwaitinfo}, requiring very little CPU time for preemption. One final detail about the alarm thread is how to wake it when additional communication is required (e.g., on thread termination). This unblocking is also done using {\tt SIGALRM}, but sent through the \code{pthread_sigqueue}. Indeed, \code{sigwait} can differentiate signals sent from \code{pthread_sigqueue} from signals sent from alarms or the kernel.
1955
1956\subsection{Scheduler}
1957Finally, an aspect that was not mentioned yet is the scheduling algorithm. Currently, the \CFA scheduler uses a single ready queue for all processors, which is the simplest approach to scheduling. Further discussion on scheduling is present in section \ref{futur:sched}.
1958
1959% ======================================================================
1960% ======================================================================
1961\section{Internal Scheduling} \label{impl:intsched}
1962% ======================================================================
1963% ======================================================================
1964The following figure is the traditional illustration of a monitor (repeated from page~\pageref{fig:ClassicalMonitor} for convenience):
1965
1966\begin{figure}[H]
1967\begin{center}
1968{\resizebox{0.4\textwidth}{!}{\input{monitor}}}
1969\end{center}
1971\end{figure}
1972
1973This picture has several components, the two most important being the entry queue and the AS-stack. The entry queue is an (almost) FIFO list where threads waiting to enter are parked, while the acceptor/signaller (AS) stack is a FILO list used for threads that have been signalled or otherwise marked as running next.
1974
1975For \CFA, this picture does not have support for blocking multiple monitors on a single condition. To support \textbf{bulk-acq} two changes to this picture are required. First, it is no longer helpful to attach the condition to \emph{a single} monitor. Secondly, the thread waiting on the condition has to be separated across multiple monitors, seen in figure \ref{fig:monitor_cfa}.
1976
1977\begin{figure}[H]
1978\begin{center}
1979{\resizebox{0.8\textwidth}{!}{\input{int_monitor}}}
1980\end{center}
1981\caption{Illustration of \CFA Monitor}
1982\label{fig:monitor_cfa}
1983\end{figure}
1984
1985This picture and the proper entry and leave algorithms (see listing \ref{lst:entry2}) is the fundamental implementation of internal scheduling. Note that when a thread is moved from the condition to the AS-stack, it is conceptually split into N pieces, where N is the number of monitors specified in the parameter list. The thread is woken up when all the pieces have popped from the AS-stacks and made active. In this picture, the threads are split into halves but this is only because there are two monitors. For a specific signalling operation every monitor needs a piece of thread on its AS-stack.
1986
1987\begin{figure}[b]
1988\begin{multicols}{2}
1989Entry
1990\begin{pseudo}
1991if monitor is free
1992        enter
1994        continue
1995else
1996        block
1997increment recursion
1998
1999\end{pseudo}
2000\columnbreak
2001Exit
2002\begin{pseudo}
2003decrement recursion
2004if recursion == 0
2005        if signal_stack not empty
2009
2010        if entry queue not empty
2012\end{pseudo}
2013\end{multicols}
2014\begin{pseudo}[caption={Entry and exit routine for monitors with internal scheduling},label={lst:entry2}]
2015\end{pseudo}
2016\end{figure}
2017
2018The solution discussed in \ref{intsched} can be seen in the exit routine of listing \ref{lst:entry2}. Basically, the solution boils down to having a separate data structure for the condition queue and the AS-stack, and unconditionally transferring ownership of the monitors but only unblocking the thread when the last monitor has transferred ownership. This solution is deadlock safe as well as preventing any potential barging. The data structures used for the AS-stack are reused extensively for external scheduling, but in the case of internal scheduling, the data is allocated using variable-length arrays on the call stack of the \code{wait} and \code{signal_block} routines.
2019
2020\begin{figure}[H]
2021\begin{center}
2022{\resizebox{0.8\textwidth}{!}{\input{monitor_structs.pstex_t}}}
2023\end{center}
2024\caption{Data structures involved in internal/external scheduling}
2025\label{fig:structs}
2026\end{figure}
2027
2028Figure \ref{fig:structs} shows a high-level representation of these data structures. The main idea behind them is that, a thread cannot contain an arbitrary number of intrusive next'' pointers for linking onto monitors. The \code{condition node} is the data structure that is queued onto a condition variable and, when signalled, the condition queue is popped and each \code{condition criterion} is moved to the AS-stack. Once all the criteria have been popped from their respective AS-stacks, the thread is woken up, which is what is shown in listing \ref{lst:entry2}.
2029
2030% ======================================================================
2031% ======================================================================
2032\section{External Scheduling}
2033% ======================================================================
2034% ======================================================================
2035Similarly to internal scheduling, external scheduling for multiple monitors relies on the idea that waiting-thread queues are no longer specific to a single monitor, as mentioned in section \ref{extsched}. For internal scheduling, these queues are part of condition variables, which are still unique for a given scheduling operation (i.e., no signal statement uses multiple conditions). However, in the case of external scheduling, there is no equivalent object which is associated with \code{waitfor} statements. This absence means the queues holding the waiting threads must be stored inside at least one of the monitors that is acquired. These monitors being the only objects that have sufficient lifetime and are available on both sides of the \code{waitfor} statement. This requires an algorithm to choose which monitor holds the relevant queue. It is also important that said algorithm be independent of the order in which users list parameters. The proposed algorithm is to fall back on monitor lock ordering (sorting by address) and specify that the monitor that is acquired first is the one with the relevant waiting queue. This assumes that the lock acquiring order is static for the lifetime of all concerned objects but that is a reasonable constraint.
2036
2037This algorithm choice has two consequences:
2038\begin{itemize}
2039        \item The queue of the monitor with the lowest address is no longer a true FIFO queue because threads can be moved to the front of the queue. These queues need to contain a set of monitors for each of the waiting threads. Therefore, another thread whose set contains the same lowest address monitor but different lower priority monitors may arrive first but enter the critical section after a thread with the correct pairing.
2040        \item The queue of the lowest priority monitor is both required and potentially unused. Indeed, since it is not known at compile time which monitor is the monitor which has the lowest address, every monitor needs to have the correct queues even though it is possible that some queues go unused for the entire duration of the program, for example if a monitor is only used in a specific pair.
2041\end{itemize}
2042Therefore, the following modifications need to be made to support external scheduling:
2043\begin{itemize}
2044        \item The threads waiting on the entry queue need to keep track of which routine they are trying to enter, and using which set of monitors. The \code{mutex} routine already has all the required information on its stack, so the thread only needs to keep a pointer to that information.
2045        \item The monitors need to keep a mask of acceptable routines. This mask contains for each acceptable routine, a routine pointer and an array of monitors to go with it. It also needs storage to keep track of which routine was accepted. Since this information is not specific to any monitor, the monitors actually contain a pointer to an integer on the stack of the waiting thread. Note that if a thread has acquired two monitors but executes a \code{waitfor} with only one monitor as a parameter, setting the mask of acceptable routines to both monitors will not cause any problems since the extra monitor will not change ownership regardless. This becomes relevant when \code{when} clauses affect the number of monitors passed to a \code{waitfor} statement.
2046        \item The entry/exit routines need to be updated as shown in listing \ref{lst:entry3}.
2047\end{itemize}
2048
2049\subsection{External Scheduling - Destructors}
2050Finally, to support the ordering inversion of destructors, the code generation needs to be modified to use a special entry routine. This routine is needed because of the storage requirements of the call order inversion. Indeed, when waiting for the destructors, storage is needed for the waiting context and the lifetime of said storage needs to outlive the waiting operation it is needed for. For regular \code{waitfor} statements, the call stack of the routine itself matches this requirement but it is no longer the case when waiting for the destructor since it is pushed on to the AS-stack for later. The \code{waitfor} semantics can then be adjusted correspondingly, as seen in listing \ref{lst:entry-dtor}
2051
2052\begin{figure}
2053\begin{multicols}{2}
2054Entry
2055\begin{pseudo}
2056if monitor is free
2057        enter
2059        continue
2061        push criteria to AS-stack
2062        continue
2063else
2064        block
2065increment recursion
2066\end{pseudo}
2067\columnbreak
2068Exit
2069\begin{pseudo}
2070decrement recursion
2071if recursion == 0
2072        if signal_stack not empty
2076                endif
2077        endif
2078
2079        if entry queue not empty
2081        endif
2082\end{pseudo}
2083\end{multicols}
2084\begin{pseudo}[caption={Entry and exit routine for monitors with internal scheduling and external scheduling},label={lst:entry3}]
2085\end{pseudo}
2086\end{figure}
2087
2088\begin{figure}
2089\begin{multicols}{2}
2090Destructor Entry
2091\begin{pseudo}
2092if monitor is free
2093        enter
2095        increment recursion
2096        return
2097create wait context
2100        push self to AS-stack
2101        baton pass
2102else
2103        wait
2104increment recursion
2105\end{pseudo}
2106\columnbreak
2107Waitfor
2108\begin{pseudo}
2110        if found destructor
2111                push destructor to AS-stack
2112                unlock all monitors
2113        else
2114                push self to AS-stack
2115                baton pass
2116        endif
2117        return
2118endif
2119if non-blocking
2120        Unlock all monitors
2121        Return
2122endif
2123
2124push self to AS-stack
2126block
2127return
2128\end{pseudo}
2129\end{multicols}
2130\begin{pseudo}[caption={Pseudo code for the \code{waitfor} routine and the \code{mutex} entry routine for destructors},label={lst:entry-dtor}]
2131\end{pseudo}
2132\end{figure}
2133
2134
2135% ======================================================================
2136% ======================================================================
2137\section{Putting It All Together}
2138% ======================================================================
2139% ======================================================================
2140
2141
2143As it was subtly alluded in section \ref{threads}, \code{thread}s in \CFA are in fact monitors, which means that all monitor features are available when using threads. For example, here is a very simple two thread pipeline that could be used for a simulator of a game engine:
2144\begin{figure}[H]
2145\begin{cfacode}[caption={Toy simulator using \code{thread}s and \code{monitor}s.},label={lst:engine-v1}]
2146// Visualization declaration
2148Frame * simulate( Simulator & this );
2149
2150// Simulation declaration
2152void render( Renderer & this );
2153
2154// Blocking call used as communication
2155void draw( Renderer & mutex this, Frame * frame );
2156
2157// Simulation loop
2158void main( Simulator & this ) {
2159        while( true ) {
2160                Frame * frame = simulate( this );
2161                draw( renderer, frame );
2162        }
2163}
2164
2165// Rendering loop
2166void main( Renderer & this ) {
2167        while( true ) {
2168                waitfor( draw, this );
2169                render( this );
2170        }
2171}
2172\end{cfacode}
2173\end{figure}
2174One of the obvious complaints of the previous code snippet (other than its toy-like simplicity) is that it does not handle exit conditions and just goes on forever. Luckily, the monitor semantics can also be used to clearly enforce a shutdown order in a concise manner:
2175\begin{figure}[H]
2176\begin{cfacode}[caption={Same toy simulator with proper termination condition.},label={lst:engine-v2}]
2177// Visualization declaration
2179Frame * simulate( Simulator & this );
2180
2181// Simulation declaration
2183void render( Renderer & this );
2184
2185// Blocking call used as communication
2186void draw( Renderer & mutex this, Frame * frame );
2187
2188// Simulation loop
2189void main( Simulator & this ) {
2190        while( true ) {
2191                Frame * frame = simulate( this );
2192                draw( renderer, frame );
2193
2194                // Exit main loop after the last frame
2195                if( frame->is_last ) break;
2196        }
2197}
2198
2199// Rendering loop
2200void main( Renderer & this ) {
2201        while( true ) {
2202                   waitfor( draw, this );
2203                or waitfor( ^?{}, this ) {
2204                        // Add an exit condition
2205                        break;
2206                }
2207
2208                render( this );
2209        }
2210}
2211
2212// Call destructor for simulator once simulator finishes
2213// Call destructor for renderer to signify shutdown
2214\end{cfacode}
2215\end{figure}
2216
2218As mentioned in section \ref{preemption}, \CFA uses preemptive threads by default but can use fibers on demand. Currently, using fibers is done by adding the following line of code to the program~:
2219\begin{cfacode}
2220unsigned int default_preemption() {
2221        return 0;
2222}
2223\end{cfacode}
2224This function is called by the kernel to fetch the default preemption rate, where 0 signifies an infinite time-slice, i.e., no preemption. However, once clusters are fully implemented, it will be possible to create fibers and \textbf{uthread} in the same system, as in listing \ref{lst:fiber-uthread}
2225\begin{figure}
2227//Cluster forward declaration
2228struct cluster;
2229
2230//Processor forward declaration
2231struct processor;
2232
2233//Construct clusters with a preemption rate
2234void ?{}(cluster& this, unsigned int rate);
2235//Construct processor and add it to cluster
2236void ?{}(processor& this, cluster& cluster);
2237//Construct thread and schedule it on cluster
2239
2240//Declare two clusters
2241cluster thread_cluster = { 10ms };                     //Preempt every 10 ms
2242cluster fibers_cluster = { 0 };                         //Never preempt
2243
2244//Construct 4 processors
2245processor processors[4] = {
2246        //2 for the thread cluster
2249        //2 for the fibers cluster
2250        fibers_cluster;
2251        fibers_cluster;
2252};
2253
2257        //Construct underlying thread to automatically
2258        //be scheduled on the thread cluster
2260}
2261
2263
2264//Declares fibers
2266void ?{}(Fiber& this) {
2267        //Construct underlying thread to automatically
2268        //be scheduled on the fiber cluster
2270}
2271
2272void main(Fiber & this);
2273\end{cfacode}
2274\end{figure}
2275
2276
2277% ======================================================================
2278% ======================================================================
2279\section{Performance Results} \label{results}
2280% ======================================================================
2281% ======================================================================
2282\section{Machine Setup}
2283Table \ref{tab:machine} shows the characteristics of the machine used to run the benchmarks. All tests were made on this machine.
2284\begin{table}[H]
2285\begin{center}
2286\begin{tabular}{| l | r | l | r |}
2287\hline
2288Architecture            & x86\_64                       & NUMA node(s)  & 8 \\
2289\hline
2290CPU op-mode(s)          & 32-bit, 64-bit                & Model name    & AMD Opteron\texttrademark  Processor 6380 \\
2291\hline
2292Byte Order                      & Little Endian                 & CPU Freq              & 2.5\si{\giga\hertz} \\
2293\hline
2294CPU(s)                  & 64                            & L1d cache     & \SI{16}{\kibi\byte} \\
2295\hline
2296Thread(s) per core      & 2                             & L1i cache     & \SI{64}{\kibi\byte} \\
2297\hline
2298Core(s) per socket      & 8                             & L2 cache              & \SI{2048}{\kibi\byte} \\
2299\hline
2300Socket(s)                       & 4                             & L3 cache              & \SI{6144}{\kibi\byte} \\
2301\hline
2302\hline
2303Operating system                & Ubuntu 16.04.3 LTS    & Kernel                & Linux 4.4-97-generic \\
2304\hline
2305Compiler                        & GCC 6.3               & Translator    & CFA 1 \\
2306\hline
2307Java version            & OpenJDK-9             & Go version    & 1.9.2 \\
2308\hline
2309\end{tabular}
2310\end{center}
2311\caption{Machine setup used for the tests}
2312\label{tab:machine}
2313\end{table}
2314
2315\section{Micro Benchmarks}
2316All benchmarks are run using the same harness to produce the results, seen as the \code{BENCH()} macro in the following examples. This macro uses the following logic to benchmark the code:
2317\begin{pseudo}
2318#define BENCH(run, result) \
2319        before = gettime(); \
2320        run; \
2321        after  = gettime(); \
2322        result = (after - before) / N;
2323\end{pseudo}
2324The method used to get time is \code{clock_gettime(CLOCK_THREAD_CPUTIME_ID);}. Each benchmark is using many iterations of a simple call to measure the cost of the call. The specific number of iterations depends on the specific benchmark.
2325
2326\subsection{Context-Switching}
2327The first interesting benchmark is to measure how long context-switches take. The simplest approach to do this is to yield on a thread, which executes a 2-step context switch. Yielding causes the thread to context-switch to the scheduler and back, more precisely: from the \textbf{uthread} to the \textbf{kthread} then from the \textbf{kthread} back to the same \textbf{uthread} (or a different one in the general case). In order to make the comparison fair, coroutines also execute a 2-step context-switch by resuming another coroutine which does nothing but suspending in a tight loop, which is a resume/suspend cycle instead of a yield. Listing \ref{lst:ctx-switch} shows the code for coroutines and threads with the results in table \ref{tab:ctx-switch}. All omitted tests are functionally identical to one of these tests. The difference between coroutines and threads can be attributed to the cost of scheduling.
2328\begin{figure}
2329\begin{multicols}{2}
2330\CFA Coroutines
2331\begin{cfacode}
2332coroutine GreatSuspender {};
2333void main(GreatSuspender& this) {
2334        while(true) { suspend(); }
2335}
2336int main() {
2337        GreatSuspender s;
2338        resume(s);
2339        BENCH(
2340                for(size_t i=0; i<n; i++) {
2341                        resume(s);
2342                },
2343                result
2344        )
2345        printf("%llu\n", result);
2346}
2347\end{cfacode}
2348\columnbreak
2350\begin{cfacode}
2351
2352
2353
2354
2355int main() {
2356
2357
2358        BENCH(
2359                for(size_t i=0; i<n; i++) {
2360                        yield();
2361                },
2362                result
2363        )
2364        printf("%llu\n", result);
2365}
2366\end{cfacode}
2367\end{multicols}
2368\begin{cfacode}[caption={\CFA benchmark code used to measure context-switches for coroutines and threads.},label={lst:ctx-switch}]
2369\end{cfacode}
2370\end{figure}
2371
2372\begin{table}
2373\begin{center}
2374\begin{tabular}{| l | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] |}
2375\cline{2-4}
2376\multicolumn{1}{c |}{} & \multicolumn{1}{c |}{ Median } &\multicolumn{1}{c |}{ Average } & \multicolumn{1}{c |}{ Standard Deviation} \\
2377\hline
2378Kernel Thread   & 241.5 & 243.86        & 5.08 \\
2379\CFA Coroutine  & 38            & 38            & 0    \\
2380\CFA Thread             & 103           & 102.96        & 2.96 \\
2381\uC Coroutine   & 46            & 45.86 & 0.35 \\
2382\uC Thread              & 98            & 99.11 & 1.42 \\
2383Goroutine               & 150           & 149.96        & 3.16 \\
2384Java Thread             & 289           & 290.68        & 8.72 \\
2385\hline
2386\end{tabular}
2387\end{center}
2388\caption{Context Switch comparison. All numbers are in nanoseconds(\si{\nano\second})}
2389\label{tab:ctx-switch}
2390\end{table}
2391
2392\subsection{Mutual-Exclusion}
2393The next interesting benchmark is to measure the overhead to enter/leave a critical-section. For monitors, the simplest approach is to measure how long it takes to enter and leave a monitor routine. Listing \ref{lst:mutex} shows the code for \CFA. To put the results in context, the cost of entering a non-inline function and the cost of acquiring and releasing a \code{pthread_mutex} lock is also measured. The results can be shown in table \ref{tab:mutex}.
2394
2395\begin{figure}
2396\begin{cfacode}[caption={\CFA benchmark code used to measure mutex routines.},label={lst:mutex}]
2397monitor M {};
2398void __attribute__((noinline)) call( M & mutex m /*, m2, m3, m4*/ ) {}
2399
2400int main() {
2401        M m/*, m2, m3, m4*/;
2402        BENCH(
2403                for(size_t i=0; i<n; i++) {
2404                        call(m/*, m2, m3, m4*/);
2405                },
2406                result
2407        )
2408        printf("%llu\n", result);
2409}
2410\end{cfacode}
2411\end{figure}
2412
2413\begin{table}
2414\begin{center}
2415\begin{tabular}{| l | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] |}
2416\cline{2-4}
2417\multicolumn{1}{c |}{} & \multicolumn{1}{c |}{ Median } &\multicolumn{1}{c |}{ Average } & \multicolumn{1}{c |}{ Standard Deviation} \\
2418\hline
2419C routine                                               & 2             & 2             & 0    \\
2420FetchAdd + FetchSub                             & 26            & 26            & 0    \\
2421Pthreads Mutex Lock                             & 31            & 31.86 & 0.99 \\
2422\uC \code{monitor} member routine               & 30            & 30            & 0    \\
2423\CFA \code{mutex} routine, 1 argument   & 41            & 41.57 & 0.9  \\
2424\CFA \code{mutex} routine, 2 argument   & 76            & 76.96 & 1.57 \\
2425\CFA \code{mutex} routine, 4 argument   & 145           & 146.68        & 3.85 \\
2426Java synchronized routine                       & 27            & 28.57 & 2.6  \\
2427\hline
2428\end{tabular}
2429\end{center}
2430\caption{Mutex routine comparison. All numbers are in nanoseconds(\si{\nano\second})}
2431\label{tab:mutex}
2432\end{table}
2433
2434\subsection{Internal Scheduling}
2435The internal-scheduling benchmark measures the cost of waiting on and signalling a condition variable. Listing \ref{lst:int-sched} shows the code for \CFA, with results table \ref{tab:int-sched}. As with all other benchmarks, all omitted tests are functionally identical to one of these tests.
2436
2437\begin{figure}
2438\begin{cfacode}[caption={Benchmark code for internal scheduling},label={lst:int-sched}]
2439volatile int go = 0;
2440condition c;
2441monitor M {};
2442M m1;
2443
2444void __attribute__((noinline)) do_call( M & mutex a1 ) { signal(c); }
2445
2447void ^?{}( T & mutex this ) {}
2448void main( T & this ) {
2449        while(go == 0) { yield(); }
2450        while(go == 1) { do_call(m1); }
2451}
2452int  __attribute__((noinline)) do_wait( M & mutex a1 ) {
2453        go = 1;
2454        BENCH(
2455                for(size_t i=0; i<n; i++) {
2456                        wait(c);
2457                },
2458                result
2459        )
2460        printf("%llu\n", result);
2461        go = 0;
2462        return 0;
2463}
2464int main() {
2465        T t;
2466        return do_wait(m1);
2467}
2468\end{cfacode}
2469\end{figure}
2470
2471\begin{table}
2472\begin{center}
2473\begin{tabular}{| l | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] |}
2474\cline{2-4}
2475\multicolumn{1}{c |}{} & \multicolumn{1}{c |}{ Median } &\multicolumn{1}{c |}{ Average } & \multicolumn{1}{c |}{ Standard Deviation} \\
2476\hline
2477Pthreads Condition Variable                     & 5902.5        & 6093.29       & 714.78 \\
2478\uC \code{signal}                                       & 322           & 323   & 3.36   \\
2479\CFA \code{signal}, 1 \code{monitor}    & 352.5 & 353.11        & 3.66   \\
2480\CFA \code{signal}, 2 \code{monitor}    & 430           & 430.29        & 8.97   \\
2481\CFA \code{signal}, 4 \code{monitor}    & 594.5 & 606.57        & 18.33  \\
2482Java \code{notify}                              & 13831.5       & 15698.21      & 4782.3 \\
2483\hline
2484\end{tabular}
2485\end{center}
2486\caption{Internal scheduling comparison. All numbers are in nanoseconds(\si{\nano\second})}
2487\label{tab:int-sched}
2488\end{table}
2489
2490\subsection{External Scheduling}
2491The Internal scheduling benchmark measures the cost of the \code{waitfor} statement (\code{_Accept} in \uC). Listing \ref{lst:ext-sched} shows the code for \CFA, with results in table \ref{tab:ext-sched}. As with all other benchmarks, all omitted tests are functionally identical to one of these tests.
2492
2493\begin{figure}
2494\begin{cfacode}[caption={Benchmark code for external scheduling},label={lst:ext-sched}]
2495volatile int go = 0;
2496monitor M {};
2497M m1;
2499
2500void __attribute__((noinline)) do_call( M & mutex a1 ) {}
2501
2502void ^?{}( T & mutex this ) {}
2503void main( T & this ) {
2504        while(go == 0) { yield(); }
2505        while(go == 1) { do_call(m1); }
2506}
2507int  __attribute__((noinline)) do_wait( M & mutex a1 ) {
2508        go = 1;
2509        BENCH(
2510                for(size_t i=0; i<n; i++) {
2511                        waitfor(call, a1);
2512                },
2513                result
2514        )
2515        printf("%llu\n", result);
2516        go = 0;
2517        return 0;
2518}
2519int main() {
2520        T t;
2521        return do_wait(m1);
2522}
2523\end{cfacode}
2524\end{figure}
2525
2526\begin{table}
2527\begin{center}
2528\begin{tabular}{| l | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] |}
2529\cline{2-4}
2530\multicolumn{1}{c |}{} & \multicolumn{1}{c |}{ Median } &\multicolumn{1}{c |}{ Average } & \multicolumn{1}{c |}{ Standard Deviation} \\
2531\hline
2532\uC \code{Accept}                                       & 350           & 350.61        & 3.11  \\
2533\CFA \code{waitfor}, 1 \code{monitor}   & 358.5 & 358.36        & 3.82  \\
2534\CFA \code{waitfor}, 2 \code{monitor}   & 422           & 426.79        & 7.95  \\
2535\CFA \code{waitfor}, 4 \code{monitor}   & 579.5 & 585.46        & 11.25 \\
2536\hline
2537\end{tabular}
2538\end{center}
2539\caption{External scheduling comparison. All numbers are in nanoseconds(\si{\nano\second})}
2540\label{tab:ext-sched}
2541\end{table}
2542
2543\subsection{Object Creation}
2544Finally, the last benchmark measures the cost of creation for concurrent objects. Listing \ref{lst:creation} shows the code for \texttt{pthread}s and \CFA threads, with results shown in table \ref{tab:creation}. As with all other benchmarks, all omitted tests are functionally identical to one of these tests. The only note here is that the call stacks of \CFA coroutines are lazily created, therefore without priming the coroutine, the creation cost is very low.
2545
2546\begin{figure}
2547\begin{center}
2549\begin{ccode}
2550int main() {
2551        BENCH(
2552                for(size_t i=0; i<n; i++) {
2555                                perror( "failure" );
2556                                return 1;
2557                        }
2558
2560                                perror( "failure" );
2561                                return 1;
2562                        }
2563                },
2564                result
2565        )
2566        printf("%llu\n", result);
2567}
2568\end{ccode}
2569
2570
2571
2573\begin{cfacode}
2574int main() {
2575        BENCH(
2576                for(size_t i=0; i<n; i++) {
2578                },
2579                result
2580        )
2581        printf("%llu\n", result);
2582}
2583\end{cfacode}
2584\end{center}
2585\begin{cfacode}[caption={Benchmark code for \texttt{pthread}s and \CFA to measure object creation},label={lst:creation}]
2586\end{cfacode}
2587\end{figure}
2588
2589\begin{table}
2590\begin{center}
2591\begin{tabular}{| l | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] | S[table-format=5.2,table-number-alignment=right] |}
2592\cline{2-4}
2593\multicolumn{1}{c |}{} & \multicolumn{1}{c |}{ Median } &\multicolumn{1}{c |}{ Average } & \multicolumn{1}{c |}{ Standard Deviation} \\
2594\hline
2595Pthreads                        & 26996 & 26984.71      & 156.6  \\
2596\CFA Coroutine Lazy     & 6             & 5.71  & 0.45   \\
2597\CFA Coroutine Eager    & 708           & 706.68        & 4.82   \\
2598\CFA Thread                     & 1173.5        & 1176.18       & 15.18  \\
2599\uC Coroutine           & 109           & 107.46        & 1.74   \\
2600\uC Thread                      & 526           & 530.89        & 9.73   \\
2601Goroutine                       & 2520.5        & 2530.93       & 61,56  \\
2602Java Thread                     & 91114.5       & 92272.79      & 961.58 \\
2603\hline
2604\end{tabular}
2605\end{center}
2606\caption{Creation comparison. All numbers are in nanoseconds(\si{\nano\second}).}
2607\label{tab:creation}
2608\end{table}
2609
2610
2611
2612\section{Conclusion}
2613This paper has achieved a minimal concurrency \textbf{api} that is simple, efficient and usable as the basis for higher-level features. The approach presented is based on a lightweight thread-system for parallelism, which sits on top of clusters of processors. This M:N model is judged to be both more efficient and allow more flexibility for users. Furthermore, this document introduces monitors as the main concurrency tool for users. This paper also offers a novel approach allowing multiple monitors to be accessed simultaneously without running into the Nested Monitor Problem~\cite{Lister77}. It also offers a full implementation of the concurrency runtime written entirely in \CFA, effectively the largest \CFA code base to date.
2614
2615
2616% ======================================================================
2617% ======================================================================
2618\section{Future Work}
2619% ======================================================================
2620% ======================================================================
2621
2622\subsection{Performance} \label{futur:perf}
2623This paper presents a first implementation of the \CFA concurrency runtime. Therefore, there is still significant work to improve performance. Many of the data structures and algorithms may change in the future to more efficient versions. For example, the number of monitors in a single \textbf{bulk-acq} is only bound by the stack size, this is probably unnecessarily generous. It may be possible that limiting the number helps increase performance. However, it is not obvious that the benefit would be significant.
2624
2625\subsection{Flexible Scheduling} \label{futur:sched}
2626An important part of concurrency is scheduling. Different scheduling algorithms can affect performance (both in terms of average and variation). However, no single scheduler is optimal for all workloads and therefore there is value in being able to change the scheduler for given programs. One solution is to offer various tweaking options to users, allowing the scheduler to be adjusted to the requirements of the workload. However, in order to be truly flexible, it would be interesting to allow users to add arbitrary data and arbitrary scheduling algorithms. For example, a web server could attach Type-of-Service information to threads and have a `ToS aware'' scheduling algorithm tailored to this specific web server. This path of flexible schedulers will be explored for \CFA.
2627
2628\subsection{Non-Blocking I/O} \label{futur:nbio}
2629While most of the parallelism tools are aimed at data parallelism and control-flow parallelism, many modern workloads are not bound on computation but on IO operations, a common case being web servers and XaaS (anything as a service). These types of workloads often require significant engineering around amortizing costs of blocking IO operations. At its core, non-blocking I/O is an operating system level feature that allows queuing IO operations (e.g., network operations) and registering for notifications instead of waiting for requests to complete. In this context, the role of the language makes Non-Blocking IO easily available and with low overhead. The current trend is to use asynchronous programming using tools like callbacks and/or futures and promises, which can be seen in frameworks like Node.js~\cite{NodeJs} for JavaScript, Spring MVC~\cite{SpringMVC} for Java and Django~\cite{Django} for Python. However, while these are valid solutions, they lead to code that is harder to read and maintain because it is much less linear.
2630
2631\subsection{Other Concurrency Tools} \label{futur:tools}
2632While monitors offer a flexible and powerful concurrent core for \CFA, other concurrency tools are also necessary for a complete multi-paradigm concurrency package. Examples of such tools can include simple locks and condition variables, futures and promises~\cite{promises}, executors and actors. These additional features are useful when monitors offer a level of abstraction that is inadequate for certain tasks.
2633
2635Simpler applications can benefit greatly from having implicit parallelism. That is, parallelism that does not rely on the user to write concurrency. This type of parallelism can be achieved both at the language level and at the library level. The canonical example of implicit parallelism is parallel for loops, which are the simplest example of a divide and conquer algorithms~\cite{uC++book}. Table \ref{lst:parfor} shows three different code examples that accomplish point-wise sums of large arrays. Note that none of these examples explicitly declare any concurrency or parallelism objects.
2636
2637\begin{table}
2638\begin{center}
2639\begin{tabular}[t]{|c|c|c|}
2640Sequential & Library Parallel & Language Parallel \\
2641\begin{cfacode}[tabsize=3]
2642void big_sum(
2643        int* a, int* b,
2644        int* o,
2645        size_t len)
2646{
2647        for(
2648                int i = 0;
2649                i < len;
2650                ++i )
2651        {
2652                o[i]=a[i]+b[i];
2653        }
2654}
2655
2656
2657
2658
2659
2660int* a[10000];
2661int* b[10000];
2662int* c[10000];
2663//... fill in a & b
2664big_sum(a,b,c,10000);
2665\end{cfacode} &\begin{cfacode}[tabsize=3]
2666void big_sum(
2667        int* a, int* b,
2668        int* o,
2669        size_t len)
2670{
2671        range ar(a, a+len);
2672        range br(b, b+len);
2673        range or(o, o+len);
2674        parfor( ai, bi, oi,
2675        [](     int* ai,
2676                int* bi,
2677                int* oi)
2678        {
2679                oi=ai+bi;
2680        });
2681}
2682
2683
2684int* a[10000];
2685int* b[10000];
2686int* c[10000];
2687//... fill in a & b
2688big_sum(a,b,c,10000);
2689\end{cfacode}&\begin{cfacode}[tabsize=3]
2690void big_sum(
2691        int* a, int* b,
2692        int* o,
2693        size_t len)
2694{
2695        parfor (ai,bi,oi)
2696            in (a, b, o )
2697        {
2698                oi = ai + bi;
2699        }
2700}
2701
2702
2703
2704
2705
2706
2707
2708int* a[10000];
2709int* b[10000];
2710int* c[10000];
2711//... fill in a & b
2712big_sum(a,b,c,10000);
2713\end{cfacode}
2714\end{tabular}
2715\end{center}
2716\caption{For loop to sum numbers: Sequential, using library parallelism and language parallelism.}
2717\label{lst:parfor}
2718\end{table}
2719
2720Implicit parallelism is a restrictive solution and therefore has its limitations. However, it is a quick and simple approach to parallelism, which may very well be sufficient for smaller applications and reduces the amount of boilerplate needed to start benefiting from parallelism in modern CPUs.
2721
2722
2723% A C K N O W L E D G E M E N T S
2724% -------------------------------
2725\section{Acknowledgements}
2726
2727Thanks to Aaron Moss, Rob Schluntz and Andrew Beach for their work on the \CFA project as well as all the discussions which helped concretize the ideas in this paper.
2728Partial funding was supplied by the Natural Sciences and Engineering Research Council of Canada and a corporate partnership with Huawei Ltd.
2729
2730
2731% B I B L I O G R A P H Y
2732% -----------------------------
2733\bibliographystyle{plain}
2734\bibliography{pl,local}
2735
2736\end{document}
2737
2738% Local Variables: %
2739% tab-width: 4 %
2740% fill-column: 120 %
2741% compile-command: "make" %
2742% End: %
Note: See TracBrowser for help on using the repository browser.