# Changes in /[9af0fe2d:ef0b456]

Ignore:
Location:
doc/theses/thierry_delisle_PhD/thesis
Files:
7 edited

Unmodified
Removed
• ## doc/theses/thierry_delisle_PhD/thesis/Makefile

 r9af0fe2d emptytree \ fairness \ io_uring \ system \ } ## Define the documents that need to be made. all: thesis.pdf thesis.pdf: ${TEXTS}${FIGURES} ${PICTURES} glossary.tex local.bib thesis.pdf:${TEXTS} ${FIGURES}${PICTURES} thesis.tex glossary.tex local.bib DOCUMENT = thesis.pdf
• ## doc/theses/thierry_delisle_PhD/thesis/local.bib

 r9af0fe2d } @manual{MAN:bsd/kqueue, title = {KQUEUE(2) - FreeBSD System Calls Manual}, url   = {https://www.freebsd.org/cgi/man.cgi?query=kqueue}, year  = {2020}, month = {may} } % Apple's MAC OS X @manual{MAN:apple/scheduler, % -------------------------------------------------- % Man Pages @manual{MAN:open, key        = "open", title      = "open(2) Linux User's Manual", year       = "2020", month      = "February", } @manual{MAN:accept, key        = "accept", title      = "accept(2) Linux User's Manual", year       = "2019", month      = "March", } @manual{MAN:select, key        = "select", title      = "select(2) Linux User's Manual", year       = "2019", month      = "March", } @manual{MAN:poll, key        = "poll", title      = "poll(2) Linux User's Manual", year       = "2019", month      = "July", } @manual{MAN:epoll, key        = "epoll", title      = "epoll(7) Linux User's Manual", year       = "2019", month      = "March", } @manual{MAN:aio, key        = "aio", title      = "aio(7) Linux User's Manual", year       = "2019", month      = "March", } @misc{MAN:io_uring, title   = {Efficient IO with io\_uring}, author  = {Axboe, Jens}, year    = "2019", month   = "March", version = {0,4}, howpublished = {\url{https://kernel.dk/io_uring.pdf}} } % -------------------------------------------------- % Wikipedia Entries @misc{wiki:taskparallel,
• ## doc/theses/thierry_delisle_PhD/thesis/text/core.tex

 r9af0fe2d \section{Design} In general, a na\"{i}ve \glsxtrshort{fifo} ready-queue does not scale with increased parallelism from \glspl{hthrd}, resulting in decreased performance. The problem is adding/removing \glspl{thrd} is a single point of contention. As shown in the evaluation sections, most production schedulers do scale when adding \glspl{hthrd}. The common solution to the single point of contention is to shard the ready-queue so each \gls{hthrd} can access the ready-queue without contention, increasing performance though lack of contention. In general, a na\"{i}ve \glsxtrshort{fifo} ready-queue does not scale with increased parallelism from \glspl{hthrd}, resulting in decreased performance. The problem is adding/removing \glspl{thrd} is a single point of contention. As shown in the evaluation sections, most production schedulers do scale when adding \glspl{hthrd}. The common solution to the single point of contention is to shard the ready-queue so each \gls{hthrd} can access the ready-queue without contention, increasing performance. \subsection{Sharding} \label{sec:sharding} An interesting approach to sharding a queue is presented in \cit{Trevors paper}. This algorithm presents a queue with a relaxed \glsxtrshort{fifo} guarantee using an array of strictly \glsxtrshort{fifo} sublists as shown in Figure~\ref{fig:base}. Each \emph{cell} of the array has a timestamp for the last operation and a pointer to a linked-list with a lock and each node in the list is marked with a timestamp indicating when it is added to the list. A push operation is done by picking a random cell, acquiring the list lock, and pushing to the list. If the cell is locked, the operation is simply retried on another random cell until a lock is acquired. A pop operation is done in a similar fashion except two random cells are picked. If both cells are unlocked with non-empty lists, the operation pops the node with the oldest cell timestamp. If one of the cells is unlocked and non-empty, the operation pops from that cell. If both cells are either locked or empty, the operation picks two new random cells and tries again. An interesting approach to sharding a queue is presented in \cit{Trevors paper}. This algorithm presents a queue with a relaxed \glsxtrshort{fifo} guarantee using an array of strictly \glsxtrshort{fifo} sublists as shown in Figure~\ref{fig:base}. Each \emph{cell} of the array has a timestamp for the last operation and a pointer to a linked-list with a lock. Each node in the list is marked with a timestamp indicating when it is added to the list. A push operation is done by picking a random cell, acquiring the list lock, and pushing to the list. If the cell is locked, the operation is simply retried on another random cell until a lock is acquired. A pop operation is done in a similar fashion except two random cells are picked. If both cells are unlocked with non-empty lists, the operation pops the node with the oldest timestamp. If one of the cells is unlocked and non-empty, the operation pops from that cell. If both cells are either locked or empty, the operation picks two new random cells and tries again. \begin{figure} \paragraph{Local Information} Figure~\ref{fig:emptytls} shows an approach using dense information, similar to the bitmap, but each \gls{hthrd} keeps its own independent copy. While this approach can offer good scalability \emph{and} low latency, the liveliness and discovery of the information can become a problem. This case is made worst in systems with few processors where even blind random picks can find \glspl{thrd} in a few tries. I built a prototype of these approaches and none of these techniques offer satisfying performance when few threads are present. All of these approach hit the same 2 problems. First, randomly picking sub-queues is very fast but means any improvement to the hit rate can easily be countered by a slow-down in look-up speed when there are empty lists. Second, the array is already as sharded to avoid contention bottlenecks, so any denser data structure tends to become a bottleneck. In all cases, these factors meant the best cases scenario, \ie many threads, would get worst throughput, and the worst-case scenario, few threads, would get a better hit rate, but an equivalent poor throughput. As a result I tried an entirely different approach. I built a prototype of these approaches and none of these techniques offer satisfying performance when few threads are present. All of these approach hit the same 2 problems. First, randomly picking sub-queues is very fast. That speed means any improvement to the hit rate can easily be countered by a slow-down in look-up speed, whether or not there are empty lists. Second, the array is already sharded to avoid contention bottlenecks, so any denser data structure tends to become a bottleneck. In all cases, these factors meant the best cases scenario, \ie many threads, would get worst throughput, and the worst-case scenario, few threads, would get a better hit rate, but an equivalent poor throughput. As a result I tried an entirely different approach. \subsection{Dynamic Entropy}\cit{https://xkcd.com/2318/} In the worst-case scenario there are only few \glspl{thrd} ready to run, or more precisely given $P$ \glspl{proc}\footnote{For simplicity, this assumes there is a one-to-one match between \glspl{proc} and \glspl{hthrd}.}, $T$ \glspl{thrd} and $\epsilon$ a very small number, than the worst case scenario can be represented by $\epsilon \ll P$, than $T = P + \epsilon$. It is important to note in this case that fairness is effectively irrelevant. Indeed, this case is close to \emph{actually matching} the model of the Ideal multi-tasking CPU'' on page \pageref{q:LinuxCFS}. In this context, it is possible to use a purely internal-locality based approach and still meet the fairness requirements. This approach simply has each \gls{proc} running a single \gls{thrd} repeatedly. Or from the shared ready-queue viewpoint, each \gls{proc} pushes to a given sub-queue and then popes from the \emph{same} subqueue. In cases where $T \gg P$, the scheduler should also achieves similar performance without affecting the fairness guarantees. In the worst-case scenario there are only few \glspl{thrd} ready to run, or more precisely given $P$ \glspl{proc}\footnote{For simplicity, this assumes there is a one-to-one match between \glspl{proc} and \glspl{hthrd}.}, $T$ \glspl{thrd} and $\epsilon$ a very small number, than the worst case scenario can be represented by $T = P + \epsilon$, with $\epsilon \ll P$. It is important to note in this case that fairness is effectively irrelevant. Indeed, this case is close to \emph{actually matching} the model of the Ideal multi-tasking CPU'' on page \pageref{q:LinuxCFS}. In this context, it is possible to use a purely internal-locality based approach and still meet the fairness requirements. This approach simply has each \gls{proc} running a single \gls{thrd} repeatedly. Or from the shared ready-queue viewpoint, each \gls{proc} pushes to a given sub-queue and then pops from the \emph{same} subqueue. The challenge is for the the scheduler to achieve good performance in both the $T = P + \epsilon$ case and the $T \gg P$ case, without affecting the fairness guarantees in the later. To handle this case, I use a pseudo random-number generator, \glsxtrshort{prng} in a novel way. When the scheduler uses a \glsxtrshort{prng} instance per \gls{proc} exclusively, the random-number seed effectively starts an encoding that produces a list of all accessed subqueues, from latest to oldest. The novel approach is to be able to replay'' the \glsxtrshort{prng} backwards and there exist \glsxtrshort{prng}s that are fast, compact \emph{and} can be run forward and backwards. Linear congruential generators~\cite{wiki:lcg} are an example of \glsxtrshort{prng}s that match these requirements. To handle this case, I use a \glsxtrshort{prng}\todo{Fix missing long form} in a novel way. There exist \glsxtrshort{prng}s that are fast, compact and can be run forward \emph{and} backwards.  Linear congruential generators~\cite{wiki:lcg} are an example of \glsxtrshort{prng}s of such \glsxtrshort{prng}s. The novel approach is to use the ability to run backwards to replay'' the \glsxtrshort{prng}. The scheduler uses an exclusive \glsxtrshort{prng} instance per \gls{proc}, the random-number seed effectively starts an encoding that produces a list of all accessed subqueues, from latest to oldest. Replaying the \glsxtrshort{prng} to identify cells accessed recently and which probably have data still cached. The algorithm works as follows:
• ## doc/theses/thierry_delisle_PhD/thesis/text/intro.tex

 r9af0fe2d While previous work on the concurrent package of \CFA focused on features and interfaces, this thesis focuses on performance, introducing \glsxtrshort{api} changes only when required by performance considerations. More specifically, this thesis concentrates on scheduling and \glsxtrshort{io}. Prior to this work, the \CFA runtime used a strictly \glsxtrshort{fifo} \gls{rQ}. This work exclusively concentrates on Linux as it's operating system since the existing \CFA runtime and compiler does not already support other operating systems. Furthermore, as \CFA is yet to be released, supporting version of Linux older that the latest version is not a goal of this work. This work exclusively concentrates on Linux as it's operating system since the existing \CFA runtime and compiler does not already support other operating systems. Furthermore, as \CFA is yet to be released, supporting version of Linux older than the latest version is not a goal of this work.
• ## doc/theses/thierry_delisle_PhD/thesis/text/io.tex

 r9af0fe2d As mentionned in Section~\ref{prev:io}, User-Level \glsxtrshort{io} requires multiplexing the \glsxtrshort{io} operations of many \glspl{thrd} onto fewer \glspl{proc} using asynchronous \glsxtrshort{io} operations. Various operating systems offer various forms of asynchronous operations and as mentioned in Chapter~\ref{intro}, this work is exclusively focuesd on Linux. \section{Existing options} Since \glsxtrshort{io} operations are generally handled by the \section{Kernel Interface} Since this work fundamentally depends on operating system support, the first step of any design is to discuss the available interfaces and pick one (or more) as the foundations of the \glsxtrshort{io} subsystem. \subsection{\lstinline|epoll|, \lstinline|poll| and \lstinline|select|} \subsection{\lstinline|O_NONBLOCK|} In Linux, files can be opened with the flag @O_NONBLOCK@~\cite{MAN:open} (or @SO_NONBLOCK@~\cite{MAN:accept}, the equivalent for sockets) to use the file descriptors in nonblocking mode''. In this mode, Neither the open() nor any subsequent I/O operations on the [opened file descriptor] will cause the calling process to wait.'' This feature can be used as the foundation for the \glsxtrshort{io} subsystem. However, for the subsystem to be able to block \glspl{thrd} until an operation completes, @O_NONBLOCK@ must be use in conjunction with a system call that monitors when a file descriptor becomes ready, \ie, the next \glsxtrshort{io} operation on it will not cause the process to wait\footnote{In this context, ready means to \emph{some} operation can be performed without blocking. It does not mean that the last operation that return \lstinline|EAGAIN| will succeed on the next try. A file that is ready to read but has only 1 byte available would be an example of this distinction.}. \subsection{Linux's AIO} There are three options to monitor file descriptors in Linux\footnote{For simplicity, this section omits to mention \lstinline|pselect| and \lstinline|ppoll|. The difference between these system calls and \lstinline|select| and \lstinline|poll| respectively is not relevant for this discussion.}, @select@~\cite{MAN:select}, @poll@~\cite{MAN:poll} and @epoll@~\cite{MAN:epoll}. All three of these options offer a system call that blocks a \gls{kthrd} until at least one of many file descriptor becomes ready. The group of file descriptors being waited on is often referred to as the \newterm{interest set}. \paragraph{\lstinline|select|} is the oldest of these options, it takes as an input a contiguous array of bits, where each bits represent a file descriptor of interest. On return, it modifies the set in place to identify which of the file descriptors changed status. This means that calling select in a loop requires re-initializing the array each time and the number of file descriptors supported has a hard limit. Another limit of @select@ is that once the call is started, the interest set can no longer be modified. Monitoring a new file descriptor generally requires aborting any in progress call to @select@\footnote{Starting a new call to \lstinline|select| in this case is possible but requires a distinct kernel thread, and as a result is not a acceptable multiplexing solution when the interest set is large and highly dynamic unless the number of parallel calls to select can be strictly bounded.}. \paragraph{\lstinline|poll|} is an improvement over select, which removes the hard limit on the number of file descriptors and the need to re-initialize the input on every call. It works using an array of structures as an input rather than an array of bits, thus allowing a more compact input for small interest sets. Like @select@, @poll@ suffers from the limitation that the interest set cannot be changed while the call is blocked. \paragraph{\lstinline|epoll|} further improves on these two functions, by allowing the interest set to be dynamically added to and removed from while a \gls{kthrd} is blocked on a call to @epoll@. This is done by creating an \emph{epoll instance} with a persistent intereset set and that is used across multiple calls. This advantage significantly reduces synchronization overhead on the part of the caller (in this case the \glsxtrshort{io} subsystem) since the interest set can be modified when adding or removing file descriptors without having to synchronize with other \glspl{kthrd} potentially calling @epoll@. However, all three of these system calls suffer from generality problems to some extent. The man page for @O_NONBLOCK@ mentions that [@O_NONBLOCK@] has no effect for regular files and block devices'', which means none of these three system calls are viable multiplexing strategies for these types of \glsxtrshort{io} operations. Furthermore, @epoll@ has been shown to have some problems with pipes and ttys\cit{Peter's examples in some fashion}. Finally, none of these are useful solutions for multiplexing \glsxtrshort{io} operations that do not have a corresponding file descriptor and can be awkward for operations using multiple file descriptors. \subsection{The POSIX asynchronous I/O (AIO)} An alternative to using @O_NONBLOCK@ is to use the AIO interface. Its interface lets programmers enqueue operations to be performed asynchronously by the kernel. Completions of these operations can be communicated in various ways, either by sending a Linux signal, spawning a new \gls{kthrd} or by polling for completion of one or more operation. For the purpose multiplexing operations, spawning a new \gls{kthrd} is counter-productive but a related solution is discussed in Section~\ref{io:morethreads}. Since using interrupts handlers can also lead to fairly complicated interactions between subsystems, I will concentrate on the different polling methods. AIO only supports read and write operations to file descriptors and those do not have the same limitation as @O_NONBLOCK@, \ie, the file descriptors can be regular files and blocked devices. It also supports batching more than one of these operations in a single system call. AIO offers two different approach to polling. @aio_error@ can be used as a spinning form of polling, returning @EINPROGRESS@ until the operation is completed, and @aio_suspend@ can be used similarly to @select@, @poll@ or @epoll@, to wait until one or more requests have completed. For the purpose of \glsxtrshort{io} multiplexing, @aio_suspend@ is the intended interface. Even if AIO requests can be submitted concurrently, @aio_suspend@ suffers from the same limitation as @select@ and @poll@, \ie, the interest set cannot be dynamically changed while a call to @aio_suspend@ is in progress. Unlike @select@ and @poll@ however, it also suffers from the limitation that it does not specify which requests have completed, meaning programmers then have to poll each request in the interest set using @aio_error@ to identify which requests have completed. This means that, like @select@ and @poll@ but not @epoll@, the time needed to examine polling results increases based in the total number of requests monitored, not the number of completed requests. AIO does not seem to be a particularly popular interface, which I believe is in part due to this less than ideal polling interface. Linus Torvalds talks about this interface as follows : \begin{displayquote} in some kind of arbitrary \textit{queue up asynchronous system call} model''. This description is actually quite close to the interface of the interface described in the next section. This description is actually quite close to the interface described in the next section. \subsection{\texttt{io\_uring}} A very recent addition to Linux, @io_uring@\cit{io\_uring} is a framework that aims to solve many of the problems listed with the above mentioned solutions. \subsection{\lstinline|io_uring|} A very recent addition to Linux, @io_uring@\cite{MAN:io_uring} is a framework that aims to solve many of the problems listed with the above mentioned interfaces. Like AIO, it represents \glsxtrshort{io} operations as entries added on a queue. But like @epoll@, new requests can be submitted while a blocking call waiting for requests to complete is already in progress. The @io_uring@ interface uses two ring buffers (referred to simply as rings) as its core, a submit ring to which programmers push \glsxtrshort{io} requests and a completion buffer which programmers poll for completion. One of the big advantages over the interfaces listed above is that it also supports a much wider range of operations. In addition to supporting reads and writes to any file descriptor like AIO, it supports other operations like @open@, @close@, @fsync@, @accept@, @connect@, @send@, @recv@, @splice@, \etc. On top of these, @io_uring@ adds many bells and whistles'' like avoiding copies between the kernel and user-space with shared memory, allowing different mechanisms to communicate with device drivers and supporting chains of requests, \ie, requests that automatically trigger followup requests on completion. \subsection{Extra Kernel Threads}\label{io:morethreads} Finally, if the operating system does not offer any satisfying forms of asynchronous \glsxtrshort{io} operations, a solution is to fake it by creating a pool of \glspl{kthrd} and delegating operations to them in order to avoid blocking \glspl{proc}. Finally, if the operating system does not offer any satisfying forms of asynchronous \glsxtrshort{io} operations, a solution is to fake it by creating a pool of \glspl{kthrd} and delegating operations to them in order to avoid blocking \glspl{proc}. The is a compromise on multiplexing. In the worst case, where all \glspl{thrd} are consistently blocking on \glsxtrshort{io}, it devolves into 1-to-1 threading. However, regardless of the frequency of \glsxtrshort{io} operations, it achieves the fundamental goal of not blocking \glspl{proc} when \glspl{thrd} are ready to run. This approach is used by languages like Go\cit{Go} and frameworks like libuv\cit{libuv}, since it has the advantage that it can easily be used across multiple operating systems. This advantage is especially relevant for languages like Go, which offer an homogenous \glsxtrshort{api} across all platforms. As opposed to C, which has a very limited standard api for \glsxtrshort{io}, \eg, the C standard library has no networking. \subsection{Discussion} These options effectively fall into two broad camps of solutions, waiting for \glsxtrshort{io} to be ready versus waiting for \glsxtrshort{io} to be completed. All operating systems that support asynchronous \glsxtrshort{io} must offer an interface along one of these lines, but the details can vary drastically. For example, Free BSD offers @kqueue@~\cite{MAN:bsd/kqueue} which behaves similarly to @epoll@ but with some small quality of life improvements, while Windows (Win32)~\cit{https://docs.microsoft.com/en-us/windows/win32/fileio/synchronous-and-asynchronous-i-o} offers overlapped I/O'' which handles submissions similarly to @O_NONBLOCK@, with extra flags on the synchronous system call, but waits for completion events, similarly to @io_uring@. For this project, I have chosen to use @io_uring@, in large parts due to its generality. While @epoll@ has been shown to be a good solution to socket \glsxtrshort{io} (\cite{DBLP:journals/pomacs/KarstenB20}), @io_uring@'s transparent support for files, pipes and more complex operations, like @splice@ and @tee@, make it a better choice as the foundation for a general \glsxtrshort{io} subsystem. \section{Event-Engine} The event engines reponsibility is to use the kernel interface to multiplex many \glsxtrshort{io} operations onto few \glspl{kthrd}. In concrete terms, this means that \glspl{thrd} enter the engine threw an interface, the event engines then starts the operation and parks the calling \glspl{thrd}, returning control to the \gls{proc}. The parked \glspl{thrd} are then rescheduled by the event engine once the desired operation has completed. \subsection{\lstinline|io_uring| in depth} Before going into details on the design of the event engine, I will present some more details on the usage of @io_uring@ which are important for the design of the engine. \begin{figure} \centering \input{io_uring.pstex_t} \caption[Overview of \lstinline|io_uring|]{Overview of \lstinline|io_uring| \smallskip\newline Two ring buffer are used to communicate with the kernel, one for completions~(right) and one for submissions~(left). The completion ring contains entries, \newterm{CQE}s: Completion Queue Entries, that are produced by the kernel when an operation completes and then consumed by the application. On the other hand, the application produces \newterm{SQE}s: Submit Queue Entries, which it appends to the submission ring for the kernel to consume. Unlike the completion ring, the submission ring does not contain the entries directly, it indexes into the SQE array (denoted \emph{S}) instead.} \label{fig:iouring} \end{figure} Figure~\ref{fig:iouring} shows an overview of @io_uring@. New \glsxtrshort{io} operations are submitted to the kernel following 4 steps which use the components shown in the figure. \paragraph{First} an @sqe@ must be allocated from the pre-allocated array (denoted \emph{S} in Figure~\ref{fig:iouring}). This array is create at the same time as the rings, is in kernel-locked memory, which means it is both visible by the kernel and the application, and has a fixed size determined at creation. How these entries are allocated is not important for the functionning of @io_uring@, the only requirement is that no entry is reused before the kernel has consumed it. \paragraph{Secondly} the @sqe@ must be filled according to the desired operation. This step is straight forward, the only detail worth mentionning is that @sqe@s have a @user_data@ field that must be filled in order to match submission and completion entries. \paragraph{Thirdly} the @sqe@ must be submitted to the submission ring, this requires appending the index of the @sqe@ to the ring following regular ring buffer steps: \lstinline|{ buffer[head] = item; head++ }|. Since the head is visible to the kernel, some memory barriers may be required to prevent the compiler to reorder these operations. Since the submission ring is a regular ring buffer, more than one @sqe@ can be added at once and the head can be updated only after the entire batch has been updated. \paragraph{Finally} the kernel must be notified of the change to the ring using the system call @io_uring_enter@. The number of elements appended to the submission ring is passed as a parameter and the number of elements consumed is returned. The @io_uring@ instance can be constructed so that this step is not required, but this requires elevated privilege and early version of @io_uring@ had additionnal restrictions. The submission side is simpler, applications call @io_uring_enter@ with the flag @IORING_ENTER_GETEVENTS@ to wait on a desired number of operations to complete. The same call can be used to both submit @sqe@s and wait for operations to complete. When operations do complete the kernel appends a @cqe@ to the completion ring and advances the head of the ring. Each @cqe@ contains the result of the operation as well as a copy of the @user_data@ field of the @sqe@ that triggered the operation. It is not necessary to call @io_uring_enter@ to get new events, the kernel can directly modify the completion ring, the system call is only needed if the application wants to block waiting on operations to complete. The @io_uring_enter@ system call is protected by a lock inside the kernel. This means that concurrent call to @io_uring_enter@ using the same instance are possible, but there is can be no performance gained from parallel calls to @io_uring_enter@. It is possible to do the first three submission steps in parallel, however, doing so requires careful synchronization. \section{Interface} Finally, the last important part of the I/O subsystem is it's interface. There are multiple approaches that can be offered to programmers, each with advantages and disadvantages. The new \glsxtrshort{io} subsystem can replace the C runtime's API or extend it. And in the later case the interface can go from very similar to vastly different. The following sections discuss some useful options using @read@ as an example. The standard Linux interface for C is : @ssize_t read(int fd, void *buf, size_t count);@. \subsection{Replacement} Replacing the C \glsxtrshort{api} \subsection{Synchronous Extension} \subsection{Asynchronous Extension} \subsection{Interface directly to \lstinline|io_uring|}
• ## doc/theses/thierry_delisle_PhD/thesis/text/runtime.tex

 r9af0fe2d \section{Clusters} \CFA allows the option to group user-level threading, in the form of clusters. Both \glspl{thrd} and \glspl{proc} belong to a specific cluster. \Glspl{thrd} are only be scheduled onto \glspl{proc} in the same cluster and scheduling is done independently of other clusters. Figure~\ref{fig:system} shows an overview of the \CFA runtime, which allows programmers to tightly control parallelism. It also opens the door to handling effects like NUMA, by pining clusters to a specific NUMA node\footnote{This is not currently implemented in \CFA, but the only hurdle left is creating a generic interface for cpu masks.}. \CFA allows the option to group user-level threading, in the form of clusters. Both \glspl{thrd} and \glspl{proc} belong to a specific cluster. \Glspl{thrd} are only scheduled onto \glspl{proc} in the same cluster and scheduling is done independently of other clusters. Figure~\ref{fig:system} shows an overview of the \CFA runtime, which allows programmers to tightly control parallelism. It also opens the door to handling effects like NUMA, by pining clusters to a specific NUMA node\footnote{This is not currently implemented in \CFA, but the only hurdle left is creating a generic interface for cpu masks.}. \begin{figure} \section{\glsxtrshort{io}}\label{prev:io} Prior to this work, the \CFA runtime did not add any particular support for \glsxtrshort{io} operations. %\CFA being built on C, this means that, While all I/O operations available in C are available in \CFA, \glsxtrshort{io} operations are designed for the POSIX threading model~\cite{pthreads}. Using these 1:1 threading operations in an M:N threading model means I/O operations block \glspl{proc} instead of \glspl{thrd}. While this can work in certain cases, it limits the number of concurrent operations to the number of \glspl{proc} rather than \glspl{thrd}. It also means deadlock can occur because all \glspl{proc} are blocked even if at least one \gls{thrd} is ready to run. A simple example of this type of deadlock would be as follows: Prior to this work, the \CFA runtime did not add any particular support for \glsxtrshort{io} operations. While all \glsxtrshort{io} operations available in C are available in \CFA, \glsxtrshort{io} operations are designed for the POSIX threading model~\cite{pthreads}. Using these 1:1 threading operations in an M:N threading model means \glsxtrshort{io} operations block \glspl{proc} instead of \glspl{thrd}. While this can work in certain cases, it limits the number of concurrent operations to the number of \glspl{proc} rather than \glspl{thrd}. It also means deadlock can occur because all \glspl{proc} are blocked even if at least one \gls{thrd} is ready to run. A simple example of this type of deadlock would be as follows: \begin{quote} Given a simple network program with 2 \glspl{thrd} and a single \gls{proc}, one \gls{thrd} sends network requests to a server and the other \gls{thrd} waits for a response from the server. If the second \gls{thrd} races ahead, it may wait for responses to requests that have not been sent yet. In theory, this should not be a problem, even if the second \gls{thrd} waits, because the first \gls{thrd} is still ready to run and should be able to get CPU time to send the request. With M:N threading, while the first \gls{thrd} is ready, the lone \gls{proc} \emph{cannot} run the first \gls{thrd} if it is blocked in the \glsxtrshort{io} operation of the second \gls{thrd}. If this happen, the system is in a synchronization deadlock\footnote{In this example, the deadlocked could be resolved if the server sends unprompted messages to the client. However, this solution is not general and may not be appropriate even in this simple case.}. \end{quote} Therefore, one of the objective of this work is to introduce \emph{User-Level \glsxtrshort{io}}, like \glslink{uthrding}{User-Level \emph{Threading}} blocks \glspl{thrd} rather than \glspl{proc} when doing \glsxtrshort{io} operations, which entails multiplexing the \glsxtrshort{io} operations of many \glspl{thrd} onto fewer \glspl{proc}. This multiplexing requires that a single \gls{proc} be able to execute multiple I/O operations in parallel. This requirement cannot be done with operations that block \glspl{proc}, \ie \glspl{kthrd}, since the first operation would prevent starting new operations for its blocking duration. Executing I/O operations in parallel requires \emph{asynchronous} \glsxtrshort{io}, sometimes referred to as \emph{non-blocking}, since the \gls{kthrd} does not block. \section{Interoperating with C} Therefore, one of the objective of this work is to introduce \emph{User-Level \glsxtrshort{io}}, like \glslink{uthrding}{User-Level \emph{Threading}} blocks \glspl{thrd} rather than \glspl{proc} when doing \glsxtrshort{io} operations, which entails multiplexing the \glsxtrshort{io} operations of many \glspl{thrd} onto fewer \glspl{proc}. This multiplexing requires that a single \gls{proc} be able to execute multiple \glsxtrshort{io} operations in parallel. This requirement cannot be done with operations that block \glspl{proc}, \ie \glspl{kthrd}, since the first operation would prevent starting new operations for its blocking duration. Executing \glsxtrshort{io} operations in parallel requires \emph{asynchronous} \glsxtrshort{io}, sometimes referred to as \emph{non-blocking}, since the \gls{kthrd} does not block. \section{Interoperating with \texttt{C}} While \glsxtrshort{io} operations are the classical example of operations that block \glspl{kthrd}, the non-blocking challenge extends to all blocking system-calls. The POSIX standard states~\cite[\S~2.9.1]{POSIX17}: \begin{quote} \begin{enumerate} \item Precisely identifying blocking C calls is difficult. \item Introducing new code can have a significant impact on general performance. \item Introducing control points code can have a significant impact on general performance. \end{enumerate} Because of these consequences, this work does not attempt to sandbox'' calls to C. Therefore, it is possible for an unidentified library calls to block a \gls{kthrd} leading to deadlocks in \CFA's M:N threading model, which would not occur in a traditional 1:1 threading model. Currently, all M:N thread systems interacting with UNIX without sandboxing suffer from this problem but manage to work very well in the majority of applications. Therefore, a complete solution to this problem is outside the scope of this thesis. Because of these consequences, this work does not attempt to sandbox'' calls to C. Therefore, it is possible calls from an unidentified library will block a \gls{kthrd} leading to deadlocks in \CFA's M:N threading model, which would not occur in a traditional 1:1 threading model. Currently, all M:N thread systems interacting with UNIX without sandboxing suffer from this problem but manage to work very well in the majority of applications. Therefore, a complete solution to this problem is outside the scope of this thesis.