Ignore:
Timestamp:
Sep 11, 2023, 12:55:43 PM (14 months ago)
Author:
caparsons <caparson@…>
Branches:
master
Children:
c8ec58e
Parents:
3ee8853
Message:

Incorporated changes in response to Trevor's comments.

Location:
doc/theses/colby_parsons_MMAth/text
Files:
7 edited

Legend:

Unmodified
Added
Removed
  • doc/theses/colby_parsons_MMAth/text/CFA_concurrency.tex

    r3ee8853 r9509d67a  
    1212\VRef[Listing]{l:cfa_thd_init} shows a typical examples of creating a \CFA user-thread type, and then as declaring processor ($N$) and thread objects ($M$).
    1313\begin{cfa}[caption={Example of \CFA user thread and processor creation},label={l:cfa_thd_init}]
    14 @thread@ my_thread {                            $\C{// user thread type (like structure}$
     14@thread@ my_thread {                            $\C{// user thread type (like structure)}$
    1515        ... // arbitrary number of field declarations
    1616};
     
    2121        @processor@ p[2];                               $\C{// add 2 processors = 3 total with starting processor}$
    2222        {
    23                 @my_thread@ t[2], * t3 = new(); $\C{// create 3 user threads, running in routine main}$
     23                @my_thread@ t[2], * t3 = new(); $\C{// create 2 stack allocated, 1 dynamic allocated user threads}$
    2424                ... // execute concurrently
    25                 delete( t3 );                           $\C{// wait for thread to end and deallocate}$
    26         } // wait for threads to end and deallocate
     25                delete( t3 );                           $\C{// wait for t3 to end and deallocate}$
     26    } // wait for threads t[0] and t[1] to end and deallocate
    2727} // deallocate additional kernel threads
    2828\end{cfa}
  • doc/theses/colby_parsons_MMAth/text/CFA_intro.tex

    r3ee8853 r9509d67a  
    99\CFA is a layer over C, is transpiled\footnote{Source to source translator.} to C, and is largely considered to be an extension of C.
    1010Beyond C, it adds productivity features, extended libraries, an advanced type-system, and many control-flow/concurrency constructions.
    11 However, \CFA stays true to the C programming style, with most code revolving around @struct@'s and routines, and respects the same rules as C.
     11However, \CFA stays true to the C programming style, with most code revolving around @struct@s and routines, and respects the same rules as C.
    1212\CFA is not object oriented as it has no notion of @this@ (receiver) and no structures with methods, but supports some object oriented ideas including constructors, destructors, and limited nominal inheritance.
    1313While \CFA is rich with interesting features, only the subset pertinent to this work is discussed here.
     
    1717References in \CFA are a layer of syntactic sugar over pointers to reduce the number of syntactic ref/deref operations needed with pointer usage.
    1818Pointers in \CFA differ from C and \CC in their use of @0p@ instead of C's @NULL@ or \CC's @nullptr@.
     19References can contain 0p in \CFA, which is the equivalent of a null reference.
    1920Examples of references are shown in \VRef[Listing]{l:cfa_ref}.
    2021
     
    6465This feature is also implemented in Pascal~\cite{Pascal}.
    6566It can exist as a stand-alone statement or wrap a routine body to expose aggregate fields.
     67If exposed fields share a name, the type system will attempt to disambiguate them based on type.
     68If the type system is unable to disambiguate the fields then the user must qualify those names to avoid a compilation error.
    6669Examples of the @with@ statement are shown in \VRef[Listing]{l:cfa_with}.
    6770
  • doc/theses/colby_parsons_MMAth/text/actors.tex

    r3ee8853 r9509d67a  
    77Actors are an indirect concurrent feature that abstracts threading away from a programmer, and instead provides \gls{actor}s and messages as building blocks for concurrency.
    88Hence, actors are in the realm of \gls{impl_concurrency}, where programmers write concurrent code without dealing with explicit thread creation or interaction.
    9 Actor message-passing is similar to channels, but with more abstraction, so there is no shared data to protect, making actors amenable in a distributed environment.
     9Actor message-passing is similar to channels, but with more abstraction, so there is no shared data to protect, making actors amenable to a distributed environment.
    1010Actors are often used for high-performance computing and other data-centric problems, where the ease of use and scalability of an actor system provides an advantage over channels.
    1111
     
    1414
    1515\section{Actor Model}
    16 The \Newterm{actor model} is a concurrent paradigm where computation is broken into units of work called actors, and the data for computation is distributed to actors in the form of messages~\cite{Hewitt73}.
     16The \Newterm{actor model} is a concurrent paradigm where an actor is used as the fundamental building-block for computation, and the data for computation is distributed to actors in the form of messages~\cite{Hewitt73}.
    1717An actor is composed of a \Newterm{mailbox} (message queue) and a set of \Newterm{behaviours} that receive from the mailbox to perform work.
    1818Actors execute asynchronously upon receiving a message and can modify their own state, make decisions, spawn more actors, and send messages to other actors.
     
    2222For example, mutual exclusion and locking are rarely relevant concepts in an actor model, as actors typically only operate on local state.
    2323
    24 An actor does not have a thread.
     24\subsection{Classic Actor System}
     25An implementation of the actor model with a theatre (group) of actors is called an \Newterm{actor system}.
     26Actor systems largely follow the actor model, but can differ in some ways.
     27
     28In an actor system, an actor does not have a thread.
    2529An actor is executed by an underlying \Newterm{executor} (kernel thread-pool) that fairly invokes each actor, where an actor invocation processes one or more messages from its mailbox.
    2630The default number of executor threads is often proportional to the number of computer cores to achieve good performance.
    2731An executor is often tunable with respect to the number of kernel threads and its scheduling algorithm, which optimize for specific actor applications and workloads \see{Section~\ref{s:ActorSystem}}.
    2832
    29 \subsection{Classic Actor System}
    30 An implementation of the actor model with a community of actors is called an \Newterm{actor system}.
    31 Actor systems largely follow the actor model, but can differ in some ways.
    3233While the semantics of message \emph{send} is asynchronous, the implementation may be synchronous or a combination.
    33 The default semantics for message \emph{receive} is \gls{fifo}, so an actor receives messages from its mailbox in temporal (arrival) order;
    34 however, messages sent among actors arrive in any order.
     34The default semantics for message \emph{receive} is \gls{fifo}, so an actor receives messages from its mailbox in temporal (arrival) order.
     35% however, messages sent among actors arrive in any order.
    3536Some actor systems provide priority-based mailboxes and/or priority-based message-selection within a mailbox, where custom message dispatchers search among or within a mailbox(es) with a predicate for specific kinds of actors and/or messages.
    36 Some actor systems provide a shared mailbox where multiple actors receive from a common mailbox~\cite{Akka}, which is contrary to the no-sharing design of the basic actor-model (and requires additional locking).
    37 For non-\gls{fifo} service, some notion of fairness (eventual progress) must exist, otherwise messages have a high latency or starve, \ie never received.
    38 Finally, some actor systems provide multiple typed-mailboxes, which then lose the actor-\lstinline{become} mechanism \see{Section~\ref{s:SafetyProductivity}}.
     37Some actor systems provide a shared mailbox where multiple actors receive from a common mailbox~\cite{Akka}, which is contrary to the no-sharing design of the basic actor-model (and may require additional locking).
     38For non-\gls{fifo} service, some notion of fairness (eventual progress) should exist, otherwise messages have a high latency or starve, \ie are never received.
     39% Finally, some actor systems provide multiple typed-mailboxes, which then lose the actor-\lstinline{become} mechanism \see{Section~\ref{s:SafetyProductivity}}.
    3940%While the definition of the actor model provides no restrictions on message ordering, actor systems tend to guarantee that messages sent from a given actor $i$ to actor $j$ arrive at actor $j$ in the order they were sent.
    4041Another way an actor system varies from the model is allowing access to shared global-state.
     
    6061Figure \ref{f:inverted_actor} shows an actor system designed as \Newterm{message-centric}, where a set of messages are scheduled and run on underlying executor threads~\cite{uC++,Nigro21}.
    6162This design is \Newterm{inverted} because actors belong to a message queue, whereas in the classic approach a message queue belongs to each actor.
    62 Now a message send must queries the actor to know which message queue to post the message.
     63Now a message send must query the actor to know which message queue to post the message to.
    6364Again, the simplest design has a single global queue of messages accessed by the executor threads, but this approach has the same contention problem by the executor threads.
    6465Therefore, the messages (mailboxes) are sharded and executor threads schedule each message, which points to its corresponding actor.
     
    176177        @actor | str_msg | int_msg;@                    $\C{// cascade sends}$
    177178        @actor | int_msg;@                                              $\C{// send}$
    178         @actor | finished_msg;@                                 $\C{// send => terminate actor (deallocation deferred)}$
     179        @actor | finished_msg;@                                 $\C{// send => terminate actor (builtin Poison-Pill)}$
    179180        stop_actor_system();                                    $\C{// waits until actors finish}\CRT$
    180181} // deallocate actor, int_msg, str_msg
     
    492493Each executor thread iterates over its own message queues until it finds one with messages.
    493494At this point, the executor thread atomically \gls{gulp}s the queue, meaning it moves the contents of message queue to a local queue of the executor thread.
     495Gulping moves the contents of the message queue as a batch rather than removing individual elements.
    494496An example of the queue gulping operation is shown in the right side of Figure \ref{f:gulp}, where an executor thread gulps queue 0 and begins to process it locally.
    495497This step allows the executor thread to process the local queue without any atomics until the next gulp.
     
    523525
    524526Since the copy queue is an array, envelopes are allocated first on the stack and then copied into the copy queue to persist until they are no longer needed.
    525 For many workloads, the copy queues grow in size to facilitate the average number of messages in flight and there are no further dynamic allocations.
     527For many workloads, the copy queues reallocate and grow in size to facilitate the average number of messages in flight and there are no further dynamic allocations.
    526528The downside of this approach is that more storage is allocated than needed, \ie each copy queue is only partially full.
    527529Comparatively, the individual envelope allocations of a list-based queue mean that the actor system always uses the minimum amount of heap space and cleans up eagerly.
     
    562564To ensure sequential actor execution and \gls{fifo} message delivery in a message-centric system, stealing requires finding and removing \emph{all} of an actor's messages, and inserting them consecutively in another message queue.
    563565This operation is $O(N)$ with a non-trivial constant.
    564 The only way for work stealing to become practical is to shard each worker's message queue, which also reduces contention, and to steal queues to eliminate queue searching.
     566The only way for work stealing to become practical is to shard each worker's message queue \see{Section~\ref{s:executor}}, which also reduces contention, and to steal queues to eliminate queue searching.
    565567
    566568Given queue stealing, the goal of the presented stealing implementation is to have an essentially zero-contention-cost stealing mechanism.
     
    576578
    577579The outline for lazy-stealing by a thief is: select a victim, scan its queues once, and return immediately if a queue is stolen.
    578 The thief then assumes it normal operation of scanning over its own queues looking for work, where stolen work is placed at the end of the scan.
     580The thief then assumes its normal operation of scanning over its own queues looking for work, where stolen work is placed at the end of the scan.
    579581Hence, only one victim is affected and there is a reasonable delay between stealing events as the thief scans its ready queue looking for its own work before potentially stealing again.
    580582This lazy examination by the thief has a low perturbation cost for victims, while still finding work in a moderately loaded system.
     
    636638% Note that a thief never exceeds its $M$/$N$ worker range because it is always exchanging queues with other workers.
    637639If no appropriate victim mailbox is found, no swap is attempted.
     640Note that since the mailbox checks happen non-atomically, the thieves effectively guess which mailbox is ripe for stealing.
     641The thief may read stale data and can end up stealing an ineligible or empty mailbox.
     642This is not a correctness issue and is addressed in Section~\ref{s:steal_prob}, but the steal will likely not be productive.
     643These unproductive steals are uncommon, but do occur with some frequency, and are a tradeoff that is made to achieve minimal victim contention.
    638644
    639645\item
     
    644650\end{enumerate}
    645651
    646 \subsection{Stealing Problem}
     652\subsection{Stealing Problem}\label{s:steal_prob}
    647653Each queue access (send or gulp) involving any worker (thief or victim) is protected using spinlock @mutex_lock@.
    648654However, to achieve the goal of almost zero contention for the victim, it is necessary that the thief does not acquire any queue spinlocks in the stealing protocol.
     
    703709None of the work-stealing actor-systems examined in this work perform well on the repeat benchmark.
    704710Hence, for all non-pathological cases, the claim is made that this stealing mechanism has a (probabilistically) zero-victim-cost in practice.
     711Future work on the work stealing system could include a backoff mechanism after failed steals to further address the pathological cases.
    705712
    706713\subsection{Queue Pointer Swap}\label{s:swap}
     
    709716The \gls{cas} is a read-modify-write instruction available on most modern architectures.
    710717It atomically compares two memory locations, and if the values are equal, it writes a new value into the first memory location.
    711 A software implementation of \gls{cas} is:
     718A sequential specification of \gls{cas} is:
    712719\begin{cfa}
    713720// assume this routine executes atomically
     
    755762
    756763Either a true memory/memory swap instruction or a \gls{dcas} would provide the ability to atomically swap two memory locations, but unfortunately neither of these instructions are supported on the architectures used in this work.
     764There are lock-free implemetions of DCAS, or more generally K-word CAS (also known as MCAS or CASN)~\cite{Harris02} and LLX/SCX~\cite{Brown14} that can be used to provide the desired atomic swap capability.
     765However, these lock-free implementations were not used as it is advantageous in the work stealing case to let an attempted atomic swap fail instead of retrying.
    757766Hence, a novel atomic swap specific to the actor use case is simulated, called \gls{qpcas}.
     767Note that this swap is \emph{not} lock-free.
    758768The \gls{qpcas} is effectively a \gls{dcas} special cased in a few ways:
    759769\begin{enumerate}
     
    766776\end{cfa}
    767777\item
    768 The values swapped are never null pointers, so a null pointer can be used as an intermediate value during the swap.
     778The values swapped are never null pointers, so a null pointer can be used as an intermediate value during the swap. As such, null is effectively used as a lock for the swap.
    769779\end{enumerate}
    770780Figure~\ref{f:qpcasImpl} shows the \CFA pseudocode for the \gls{qpcas}.
     
    862872The concurrent proof of correctness is shown through the existence of an invariant.
    863873The invariant states when a queue pointer is set to @0p@ by a thief, then the next write to the pointer can only be performed by the same thief.
     874This is effictively a mutual exclusion condition for the later write.
    864875To show that this invariant holds, it is shown that it is true at each step of the swap.
    865876\begin{itemize}
     
    10111022The intuition behind this heuristic is that the slowest worker receives help via work stealing until it becomes a thief, which indicates that it has caught up to the pace of the rest of the workers.
    10121023This heuristic should ideally result in lowered latency for message sends to victim workers that are overloaded with work.
     1024It must be acknowledged that this linear search could cause a lot of cache coherence traffic.
     1025Future work on this heuristic could include introducing a search that has less impact on caching.
    10131026A negative side-effect of this heuristic is that if multiple thieves steal at the same time, they likely steal from the same victim, which increases the chance of contention.
    10141027However, given that workers have multiple queues, often in the tens or hundreds of queues, it is rare for two thieves to attempt stealing from the same queue.
     
    10281041\CFA's actor system comes with a suite of safety and productivity features.
    10291042Most of these features are only present in \CFA's debug mode, and hence, have zero-cost in no-debug mode.
    1030 The suit of features include the following.
     1043The suite of features include the following.
    10311044\begin{itemize}
    10321045\item Static-typed message sends:
    1033 If an actor does not support receiving a given message type, the receive call is rejected at compile time, allowing unsupported messages to never be sent to an actor.
     1046If an actor does not support receiving a given message type, the receive call is rejected at compile time, preventing unsupported messages from being sent to an actor.
    10341047
    10351048\item Detection of message sends to Finished/Destroyed/Deleted actors:
     
    10421055
    10431056\item When an executor is configured, $M >= N$.
    1044 That is, each worker must receive at least one mailbox queue, otherwise the worker spins and never does any work.
     1057That is, each worker must receive at least one mailbox queue, since otherwise a worker cannot receive any work without a queue pull messages from.
    10451058
    10461059\item Detection of unsent messages:
     
    11001113\begin{list}{\arabic{enumi}.}{\usecounter{enumi}\topsep=5pt\parsep=5pt\itemsep=0pt}
    11011114\item
    1102 Supermicro SYS--6029U--TR4 Intel Xeon Gold 5220R 24--core socket, hyper-threading $\times$ 2 sockets (48 process\-ing units) 2.2GHz, running Linux v5.8.0--59--generic
    1103 \item
    1104 Supermicro AS--1123US--TR4 AMD EPYC 7662 64--core socket, hyper-threading $\times$ 2 sockets (256 processing units) 2.0 GHz, running Linux v5.8.0--55--generic
     1115Supermicro SYS--6029U--TR4 Intel Xeon Gold 5220R 24--core socket, hyper-threading $\times$ 2 sockets (96 process\-ing units), running Linux v5.8.0--59--generic
     1116\item
     1117Supermicro AS--1123US--TR4 AMD EPYC 7662 64--core socket, hyper-threading $\times$ 2 sockets (256 processing units), running Linux v5.8.0--55--generic
    11051118\end{list}
    11061119
     
    11121125All benchmarks are run 5 times and the median is taken.
    11131126Error bars showing the 95\% confidence intervals appear on each point in the graphs.
     1127The confidence intervals are calculated using bootstrapping to avoid normality assumptions.
    11141128If the confidence bars are small enough, they may be obscured by the data point.
    11151129In this section, \uC is compared to \CFA frequently, as the actor system in \CFA is heavily based off of \uC's actor system.
  • doc/theses/colby_parsons_MMAth/text/channels.tex

    r3ee8853 r9509d67a  
    2020Neither Go nor \CFA channels have the restrictions of the early channel-based concurrent systems.
    2121
    22 Other popular languages and libraries that provide channels include C++ Boost~\cite{boost:channel}, Rust~\cite{rust:channel}, Haskell~\cite{haskell:channel}, and OCaml~\cite{ocaml:channel}.
     22Other popular languages and libraries that provide channels include C++ Boost~\cite{boost:channel}, Rust~\cite{rust:channel}, Haskell~\cite{haskell:channel}, OCaml~\cite{ocaml:channel}, and Kotlin~\cite{kotlin:channel}.
    2323Boost channels only support asynchronous (non-blocking) operations, and Rust channels are limited to only having one consumer per channel.
    2424Haskell channels are unbounded in size, and OCaml channels are zero-size.
    2525These restrictions in Haskell and OCaml are likely due to their functional approach, which results in them both using a list as the underlying data structure for their channel.
    2626These languages and libraries are not discussed further, as their channel implementation is not comparable to the bounded-buffer style channels present in Go and \CFA.
     27Kotlin channels are comparable to Go and \CFA, but unfortunately they were not identified as a comparator until after presentation of this thesis and are omitted due to time constraints.
    2728
    2829\section{Producer-Consumer Problem}
     
    3132In the problem, threads interact with a buffer in two ways: producing threads insert values into the buffer and consuming threads remove values from the buffer.
    3233In general, a buffer needs protection to ensure a producer only inserts into a non-full buffer and a consumer only removes from a non-empty buffer (synchronization).
    33 As well, a buffer needs protection from concurrent access by multiple producers or consumers attempting to insert or remove simultaneously (MX).
     34As well, a buffer needs protection from concurrent access by multiple producers or consumers attempting to insert or remove simultaneously, which is often provided by MX.
    3435
    3536\section{Channel Size}\label{s:ChannelSize}
     
    4142Fixed sized (bounded) implies the communication is mostly asynchronous, \ie the producer can proceed up to the buffer size and vice versa for the consumer with respect to removal, at which point the producer/consumer would wait.
    4243\item
    43 Infinite sized (unbounded) implies the communication is asynchronous, \ie the producer never waits but the consumer waits when the buffer is empty.
    44 Since memory is finite, all unbounded buffers are ultimately bounded;
    45 this restriction must be part of its implementation.
     44Infinite sized (unbounded) implies the communication is asymmetrically asynchronous, \ie the producer never waits but the consumer waits when the buffer is empty.
    4645\end{enumerate}
    4746
     
    5049However, like MX, a buffer should ensure every value is eventually removed after some reasonable bounded time (no long-term starvation).
    5150The simplest way to prevent starvation is to implement the buffer as a queue, either with a cyclic array or linked nodes.
     51While \gls{fifo} is not required for producer-consumer problem correctness, it is a desired property in channels as it provides predictable and often relied upon channel ordering behaviour to users.
    5252
    5353\section{First-Come First-Served}
    54 As pointed out, a bounded buffer requires MX among multiple producers or consumers.
     54As pointed out, a bounded buffer implementation often provides MX among multiple producers or consumers.
    5555This MX should be fair among threads, independent of the \gls{fifo} buffer being fair among values.
    5656Fairness among threads is called \gls{fcfs} and was defined by Lamport~\cite[p.~454]{Lamport74}.
     
    6666
    6767\section{Channel Implementation}\label{s:chan_impl}
    68 Currently, only the Go and Erlang programming languages provide user-level threading where the primary communication mechanism is channels.
    69 Both Go and Erlang have user-level threading and preemptive scheduling, and both use channels for communication.
    70 Go provides multiple homogeneous channels; each have a single associated type.
     68The programming languages Go, Kotlin, and Erlang provide user-level threading where the primary communication mechanism is channels.
     69These languages have user-level threading and preemptive scheduling, and both use channels for communication.
     70Go and Kotlin provide multiple homogeneous channels; each have a single associated type.
    7171Erlang, which is closely related to actor systems, provides one heterogeneous channel per thread (mailbox) with a typed receive pattern.
    72 Go encourages users to communicate via channels, but provides them as an optional language feature.
     72Go and Kotlin encourage users to communicate via channels, but provides them as an optional language feature.
    7373On the other hand, Erlang's single heterogeneous channel is a fundamental part of the threading system design; using it is unavoidable.
    74 Similar to Go, \CFA's channels are offered as an optional language feature.
     74Similar to Go and Kotlin, \CFA's channels are offered as an optional language feature.
    7575
    7676While iterating on channel implementation, experiments were conducted that varied the producer-consumer algorithm and lock type used inside the channel.
     
    8383The Go channel implementation utilizes cooperation among threads to achieve good performance~\cite{go:chan}.
    8484This cooperation only occurs when producers or consumers need to block due to the buffer being full or empty.
     85After a producer blocks it must wait for a consumer to signal it and vice versa.
     86The consumer or producer that signals a blocked thread is called the signalling thread.
    8587In these cases, a blocking thread stores their relevant data in a shared location and the signalling thread completes the blocking thread's operation before waking them;
    8688\ie the blocking thread has no work to perform after it unblocks because the signalling threads has done this work.
     
    8890First, each thread interacting with the channel only acquires and releases the internal channel lock once.
    8991As a result, contention on the internal lock is decreased; only entering threads compete for the lock since unblocking threads do not reacquire the lock.
    90 The other advantage of Go's wait-morphing approach is that it eliminates the bottleneck of waiting for signalled threads to run.
     92The other advantage of Go's wait-morphing approach is that it eliminates the need to wait for signalled threads to run.
    9193Note that the property of acquiring/releasing the lock only once can also be achieved with a different form of cooperation, called \Newterm{baton passing}.
    9294Baton passing occurs when one thread acquires a lock but does not release it, and instead signals a thread inside the critical section, conceptually ``passing'' the mutual exclusion from the signalling thread to the signalled thread.
     
    9496the wait-morphing approach has threads cooperate by completing the signalled thread's operation, thus removing a signalled thread's need for mutual exclusion after unblocking.
    9597While baton passing is useful in some algorithms, it results in worse channel performance than the Go approach.
    96 In the baton-passing approach, all threads need to wait for the signalled thread to reach the front of the ready queue, context switch, and run before other operations on the channel can proceed, since the signalled thread holds mutual exclusion;
     98In the baton-passing approach, all threads need to wait for the signalled thread to unblock and run before other operations on the channel can proceed, since the signalled thread holds mutual exclusion;
    9799in the wait-morphing approach, since the operation is completed before the signal, other threads can continue to operate on the channel without waiting for the signalled thread to run.
    98100
  • doc/theses/colby_parsons_MMAth/text/conclusion.tex

    r3ee8853 r9509d67a  
    2020\item Channels with comparable performance to Go, which have safety and productivity features including deadlock detection and an easy-to-use exception-based channel @close@ routine.
    2121\item An in-memory actor system, which achieves the lowest latency message send of systems tested due to the novel copy-queue data structure.
    22 \item As well, the actor system has built-in detection of six common actor errors, with excellent performance compared to other systems across all benchmarks.
    23 \item A @waituntil@ statement, which tackles the hard problem of allowing a thread to safely wait synchronously for an arbitrary set of concurrent resources.
     22\item As well, the actor system has built-in detection of six common actor errors, with excellent performance compared to other systems across all benchmarks presented in this thesis.
     23\item A @waituntil@ statement, which tackles the hard problem of allowing a thread wait synchronously for an arbitrary set of concurrent resources.
    2424\end{enumerate}
    2525
  • doc/theses/colby_parsons_MMAth/text/mutex_stmt.tex

    r3ee8853 r9509d67a  
    8383\end{figure}
    8484
    85 Like Java, \CFA monitors have \Newterm{multi-acquire} semantics so the thread in the monitor may acquire it multiple times without deadlock, allowing recursion and calling of other MX functions.
     85Like Java, \CFA monitors have \Newterm{multi-acquire} (reentrant locking) semantics so the thread in the monitor may acquire it multiple times without deadlock, allowing recursion and calling of other MX functions.
    8686For robustness, \CFA monitors ensure the monitor lock is released regardless of how an acquiring function ends, normal or exceptional, and returning a shared variable is safe via copying before the lock is released.
    8787Monitor objects can be passed through multiple helper functions without acquiring mutual exclusion, until a designated function associated with the object is called.
     
    104104}
    105105\end{cfa}
    106 The \CFA monitor implementation ensures multi-lock acquisition is done in a deadlock-free manner regardless of the number of MX parameters and monitor arguments. It it important to note that \CFA monitors do not attempt to solve the nested monitor problem~\cite{Lister77}.
     106The \CFA monitor implementation ensures multi-lock acquisition is done in a deadlock-free manner regardless of the number of MX parameters and monitor arguments via resource ordering.
     107It it important to note that \CFA monitors do not attempt to solve the nested monitor problem~\cite{Lister77}.
    107108
    108109\section{\lstinline{mutex} statement}
     
    165166In detail, the mutex statement has a clause and statement block, similar to a conditional or loop statement.
    166167The clause accepts any number of lockable objects (like a \CFA MX function prototype), and locks them for the duration of the statement.
    167 The locks are acquired in a deadlock free manner and released regardless of how control-flow exits the statement.
     168The locks are acquired in a deadlock-free manner and released regardless of how control-flow exits the statement.
     169Note that this deadlock-freedom has some limitations \see{\VRef{s:DeadlockAvoidance}}.
    168170The mutex statement provides easy lock usage in the common case of lexically wrapping a CS.
    169171Examples of \CFA mutex statement are shown in \VRef[Listing]{l:cfa_mutex_ex}.
     
    210212Like Java, \CFA introduces a new statement rather than building from existing language features, although \CFA has sufficient language features to mimic \CC RAII locking.
    211213This syntactic choice makes MX explicit rather than implicit via object declarations.
    212 Hence, it is easier for programmers and language tools to identify MX points in a program, \eg scan for all @mutex@ parameters and statements in a body of code.
     214Hence, it is easy for programmers and language tools to identify MX points in a program, \eg scan for all @mutex@ parameters and statements in a body of code; similar scanning can be done with Java's @synchronized@.
    213215Furthermore, concurrent safety is provided across an entire program for the complex operation of acquiring multiple locks in a deadlock-free manner.
    214216Unlike Java, \CFA's mutex statement and \CC's @scoped_lock@ both use parametric polymorphism to allow user defined types to work with this feature.
     
    231233thread$\(_2\)$ : sout | "uvw" | "xyz";
    232234\end{cfa}
    233 any of the outputs can appear, included a segment fault due to I/O buffer corruption:
     235any of the outputs can appear:
    234236\begin{cquote}
    235237\small\tt
     
    260262mutex( sout ) { // acquire stream lock for sout for block duration
    261263        sout | "abc";
    262         mutex( sout ) sout | "uvw" | "xyz"; // OK because sout lock is recursive
     264        sout | "uvw" | "xyz";
    263265        sout | "def";
    264266} // implicitly release sout lock
    265267\end{cfa}
    266 The inner lock acquire is likely to occur through a function call that does a thread-safe print.
    267268
    268269\section{Deadlock Avoidance}\label{s:DeadlockAvoidance}
     
    309310For fewer than 7 locks ($2^3-1$), the sort is unrolled performing the minimum number of compare and swaps for the given number of locks;
    310311for 7 or more locks, insertion sort is used.
    311 Since it is extremely rare to hold more than 6 locks at a time, the algorithm is fast and executes in $O(1)$ time.
    312 Furthermore, lock addresses are unique across program execution, even for dynamically allocated locks, so the algorithm is safe across the entire program execution.
     312It is assumed to be rare to hold more than 6 locks at a time.
     313For 6 or fewer locks the algorithm is fast and executes in $O(1)$ time.
     314Furthermore, lock addresses are unique across program execution, even for dynamically allocated locks, so the algorithm is safe across the entire program execution, as long as lifetimes of objects are appropriately managed.
     315For example, deleting a lock and allocating another one could give the new lock the same address as the deleted one, however deleting a lock in use by another thread is a programming error irrespective of the usage of the @mutex@ statement.
    313316
    314317The downside to the sorting approach is that it is not fully compatible with manual usages of the same locks outside the @mutex@ statement, \ie the lock are acquired without using the @mutex@ statement.
     
    338341\end{cquote}
    339342Comparatively, if the @scoped_lock@ is used and the same locks are acquired elsewhere, there is no concern of the @scoped_lock@ deadlocking, due to its avoidance scheme, but it may livelock.
    340 The convenience and safety of the @mutex@ statement, \ie guaranteed lock release with exceptions, should encourage programmers to always use it for locking, mitigating any deadlock scenario versus combining manual locking with the mutex statement.
     343The convenience and safety of the @mutex@ statement, \ie guaranteed lock release with exceptions, should encourage programmers to always use it for locking, mitigating most deadlock scenarios versus combining manual locking with the mutex statement.
    341344Both \CC and the \CFA do not provide any deadlock guarantees for nested @scoped_lock@s or @mutex@ statements.
    342345To do so would require solving the nested monitor problem~\cite{Lister77}, which currently does not have any practical solutions.
     
    344347\section{Performance}
    345348Given the two multi-acquisition algorithms in \CC and \CFA, each with differing advantages and disadvantages, it interesting to compare their performance.
    346 Comparison with Java is not possible, since it only takes a single lock.
     349Comparison with Java was not conducted, since the synchronized statement only takes a single object and does not provide deadlock avoidance or prevention.
    347350
    348351The comparison starts with a baseline that acquires the locks directly without a mutex statement or @scoped_lock@ in a fixed ordering and then releases them.
     
    356359Each variation is run 11 times on 2, 4, 8, 16, 24, 32 cores and with 2, 4, and 8 locks being acquired.
    357360The median is calculated and is plotted alongside the 95\% confidence intervals for each point.
     361The confidence intervals are calculated using bootstrapping to avoid normality assumptions.
    358362
    359363\begin{figure}
     
    388392}
    389393\end{cfa}
    390 \caption{Deadlock avoidance benchmark pseudocode}
     394\caption{Deadlock avoidance benchmark \CFA pseudocode}
    391395\label{l:deadlock_avoid_pseudo}
    392396\end{figure}
     
    396400% sudo dmidecode -t system
    397401\item
    398 Supermicro AS--1123US--TR4 AMD EPYC 7662 64--core socket, hyper-threading $\times$ 2 sockets (256 processing units) 2.0 GHz, TSO memory model, running Linux v5.8.0--55--generic, gcc--10 compiler
     402Supermicro AS--1123US--TR4 AMD EPYC 7662 64--core socket, hyper-threading $\times$ 2 sockets (256 processing units), TSO memory model, running Linux v5.8.0--55--generic, gcc--10 compiler
    399403\item
    400 Supermicro SYS--6029U--TR4 Intel Xeon Gold 5220R 24--core socket, hyper-threading $\times$ 2 sockets (48 processing units) 2.2GHz, TSO memory model, running Linux v5.8.0--59--generic, gcc--10 compiler
     404Supermicro SYS--6029U--TR4 Intel Xeon Gold 5220R 24--core socket, hyper-threading $\times$ 2 sockets (96 processing units), TSO memory model, running Linux v5.8.0--59--generic, gcc--10 compiler
    401405\end{list}
    402406%The hardware architectures are different in threading (multithreading vs hyper), cache structure (MESI or MESIF), NUMA layout (QPI vs HyperTransport), memory model (TSO vs WO), and energy/thermal mechanisms (turbo-boost).
     
    411415For example, on the AMD machine with 32 threads and 8 locks, the benchmarks would occasionally livelock indefinitely, with no threads making any progress for 3 hours before the experiment was terminated manually.
    412416It is likely that shorter bouts of livelock occurred in many of the experiments, which would explain large confidence intervals for some of the data points in the \CC data.
    413 In Figures~\ref{f:mutex_bench8_AMD} and \ref{f:mutex_bench8_Intel} there is the counter-intuitive result of the mutex statement performing better than the baseline.
     417In Figures~\ref{f:mutex_bench8_AMD} and \ref{f:mutex_bench8_Intel} there is the counter-intuitive result of the @mutex@ statement performing better than the baseline.
    414418At 7 locks and above the mutex statement switches from a hard coded sort to insertion sort, which should decrease performance.
    415419The hard coded sort is branch-free and constant-time and was verified to be faster than insertion sort for 6 or fewer locks.
    416 It is likely the increase in throughput compared to baseline is due to the delay spent in the insertion sort, which decreases contention on the locks.
    417 
     420Part of the difference in throughput compared to baseline is due to the delay spent in the insertion sort, which decreases contention on the locks.
     421This was verified to be part of the difference in throughput by experimenting with varying NCS delays in the baseline; however it only comprises a small portion of difference.
     422It is possible that the baseline is slowed down or the @mutex@ is sped up by other factors that are not easily identifiable.
    418423
    419424\begin{figure}
  • doc/theses/colby_parsons_MMAth/text/waituntil.tex

    r3ee8853 r9509d67a  
    168168Go's @select@ has the same exclusive-or semantics as the ALT primitive from Occam and associated code blocks for each clause like ALT and Ada.
    169169However, unlike Ada and ALT, Go does not provide guards for the \lstinline[language=go]{case} clauses of the \lstinline[language=go]{select}.
    170 As such, the exponential blowup can be seen comparing Go and \uC in Figure~\label{f:AdaMultiplexing}.
     170As such, the exponential blowup can be seen comparing Go and \uC in Figure~\ref{f:AdaMultiplexing}.
    171171Go also provides a timeout via a channel and a @default@ clause like Ada @else@ for asynchronous multiplexing.
    172172
     
    519519In following example, either channel @C1@ or @C2@ must be satisfied but nothing can be done for at least 1 or 3 seconds after the channel read, respectively.
    520520\begin{cfa}[deletekeywords={timeout}]
    521 waituntil( i << C1 ); and waituntil( timeout( 1`s ) );
    522 or waituntil( i << C2 ); and waituntil( timeout( 3`s ) );
     521waituntil( i << C1 ){} and waituntil( timeout( 1`s ) ){}
     522or waituntil( i << C2 ){} and waituntil( timeout( 3`s ) ){}
    523523\end{cfa}
    524524If only @C2@ is satisfied, \emph{both} timeout code-blocks trigger because 1 second occurs before 3 seconds.
     
    542542Now the unblocked WUT is guaranteed to have a satisfied resource and its code block can safely executed.
    543543The insertion circumvents the channel buffer via the wait-morphing in the \CFA channel implementation \see{Section~\ref{s:chan_impl}}, allowing @waituntil@ channel unblocking to not be special-cased.
     544Note that all channel operations are fair and no preference is given between @waituntil@ and direct channel operations when unblocking.
    544545
    545546Furthermore, if both @and@ and @or@ operators are used, the @or@ operations stop behaving like exclusive-or due to the race among channel operations, \eg:
Note: See TracChangeset for help on using the changeset viewer.