Ignore:
Timestamp:
Apr 19, 2022, 3:00:04 PM (4 years ago)
Author:
m3zulfiq <m3zulfiq@…>
Branches:
ADT, ast-experimental, master, pthread-emulation, qualifiedEnum, stuck-waitfor-destruct
Children:
5b84a321
Parents:
ba897d21 (diff), bb7c77d (diff)
Note: this is a merge changeset, the changes displayed below correspond to the merge itself.
Use the (diff) links above to see all the changes relative to each parent.
Message:

added benchmark and evaluations chapter to thesis

Location:
doc/theses/thierry_delisle_PhD/thesis
Files:
4 added
10 edited

Legend:

Unmodified
Added
Removed

  • doc/theses/thierry_delisle_PhD/thesis/Makefile

    rba897d21 r2e9b59b  
    2929PICTURES = ${addsuffix .pstex, \
    3030        base \
     31        base_avg \
     32        cache-share \
     33        cache-noshare \
    3134        empty \
    3235        emptybit \
    … …  
    3841        system \
    3942        cycle \
     43        result.cycle.jax.ops \
    4044}
    4145
    … …  
    112116        python3 $< $@
    113117
     118build/result.%.ns.svg : data/% | ${Build}
     119        ../../../../benchmark/plot.py -f $< -o $@ -y "ns per ops"
     120
     121build/result.%.ops.svg : data/% | ${Build}
     122        ../../../../benchmark/plot.py -f $< -o $@ -y "Ops per second"
     123
    114124## pstex with inverted colors
    115125%.dark.pstex : fig/%.fig Makefile | ${Build}

  • doc/theses/thierry_delisle_PhD/thesis/fig/base.fig

    rba897d21 r2e9b59b  
    8989         5700 5210 5550 4950 5250 4950 5100 5210 5250 5470 5550 5470
    9090         5700 5210
     912 1 1 1 0 7 50 -1 -1 4.000 0 0 -1 0 0 2
     92         3600 5700 3600 1200
     932 1 1 1 0 7 50 -1 -1 4.000 0 0 -1 0 0 2
     94         4800 5700 4800 1200
     952 1 1 1 0 7 50 -1 -1 4.000 0 0 -1 0 0 2
     96         6000 5700 6000 1200
    91974 2 -1 50 -1 0 12 0.0000 2 135 630 2100 3075 Threads\001
    92984 2 -1 50 -1 0 12 0.0000 2 165 450 2100 2850 Ready\001

  • doc/theses/thierry_delisle_PhD/thesis/glossary.tex

    rba897d21 r2e9b59b  
    101101
    102102\longnewglossaryentry{at}
    103 {name={fred}}
     103{name={task}}
    104104{
    105105Abstract object representing an unit of work. Systems will offer one or more concrete implementations of this concept (\eg \gls{kthrd}, \gls{job}), however, most of the concept of schedulings are independent of the particular implementations of the work representation. For this reason, this document use the term \Gls{at} to mean any representation and not one in particular.

  • doc/theses/thierry_delisle_PhD/thesis/local.bib

    rba897d21 r2e9b59b  
    685685  note = "[Online; accessed 9-February-2021]"
    686686}
     687
     688@misc{wiki:rcu,
     689  author = "{Wikipedia contributors}",
     690  title = "Read-copy-update --- {W}ikipedia{,} The Free Encyclopedia",
     691  year = "2022",
     692  url = "https://en.wikipedia.org/wiki/Linear_congruential_generator",
     693  note = "[Online; accessed 12-April-2022]"
     694}
     695
     696@misc{wiki:rwlock,
     697  author = "{Wikipedia contributors}",
     698  title = "Readers-writer lock --- {W}ikipedia{,} The Free Encyclopedia",
     699  year = "2021",
     700  url = "https://en.wikipedia.org/wiki/Readers%E2%80%93writer_lock",
     701  note = "[Online; accessed 12-April-2022]"
     702}

  • doc/theses/thierry_delisle_PhD/thesis/text/core.tex

    rba897d21 r2e9b59b  
    33Before discussing scheduling in general, where it is important to address systems that are changing states, this document discusses scheduling in a somewhat ideal scenario, where the system has reached a steady state. For this purpose, a steady state is loosely defined as a state where there are always \glspl{thrd} ready to run and the system has the resources necessary to accomplish the work, \eg, enough workers. In short, the system is neither overloaded nor underloaded.
    44
    5 I believe it is important to discuss the steady state first because it is the easiest case to handle and, relatedly, the case in which the best performance is to be expected. As such, when the system is either overloaded or underloaded, a common approach is to try to adapt the system to this new load and return to the steady state, \eg, by adding or removing workers. Therefore, flaws in scheduling the steady state can to be pervasive in all states.
     5It is important to discuss the steady state first because it is the easiest case to handle and, relatedly, the case in which the best performance is to be expected. As such, when the system is either overloaded or underloaded, a common approach is to try to adapt the system to this new load and return to the steady state, \eg, by adding or removing workers. Therefore, flaws in scheduling the steady state tend to be pervasive in all states.
    66
    77\section{Design Goals}
    … …  
    2525It is important to note that these guarantees are expected only up to a point. \Glspl{thrd} that are ready to run should not be prevented to do so, but they still share the limited hardware resources. Therefore, the guarantee is considered respected if a \gls{thrd} gets access to a \emph{fair share} of the hardware resources, even if that share is very small.
    2626
    27 Similarly the performance guarantee, the lack of interference among threads, is only relevant up to a point. Ideally, the cost of running and blocking should be constant regardless of contention, but the guarantee is considered satisfied if the cost is not \emph{too high} with or without contention. How much is an acceptable cost is obviously highly variable. For this document, the performance experimentation attempts to show the cost of scheduling is at worst equivalent to existing algorithms used in popular languages. This demonstration can be made by comparing applications built in \CFA to applications built with other languages or other models. Recall programmer expectation is that the impact of the scheduler can be ignored. Therefore, if the cost of scheduling is equivalent to or lower than other popular languages, I consider the guarantee achieved.
     27Similarly the performance guarantee, the lack of interference among threads, is only relevant up to a point. Ideally, the cost of running and blocking should be constant regardless of contention, but the guarantee is considered satisfied if the cost is not \emph{too high} with or without contention. How much is an acceptable cost is obviously highly variable. For this document, the performance experimentation attempts to show the cost of scheduling is at worst equivalent to existing algorithms used in popular languages. This demonstration can be made by comparing applications built in \CFA to applications built with other languages or other models. Recall programmer expectation is that the impact of the scheduler can be ignored. Therefore, if the cost of scheduling is compatitive to other popular languages, the guarantee will be consider achieved.
    2828
    2929More precisely the scheduler should be:
    … …  
    3333\end{itemize}
    3434
    35 \subsection{Fairness vs Scheduler Locality}
     35\subsection{Fairness Goals}
     36For this work fairness will be considered as having two strongly related requirements: true starvation freedom and ``fast'' load balancing.
     37
     38\paragraph{True starvation freedom} is more easily defined: As long as at least one \proc continues to dequeue \ats, all read \ats should be able to run eventually.
     39In any running system, \procs can stop dequeing \ats if they start running a \at that will simply never park.
     40Traditional workstealing schedulers do not have starvation freedom in these cases.
     41Now this requirement begs the question, what about preemption?
     42Generally speaking preemption happens on the timescale of several milliseconds, which brings us to the next requirement: ``fast'' load balancing.
     43
     44\paragraph{Fast load balancing} means that load balancing should happen faster than preemption would normally allow.
     45For interactive applications that need to run at 60, 90, 120 frames per second, \ats having to wait for several millseconds to run are effectively starved.
     46Therefore load-balancing should be done at a faster pace, one that can detect starvation at the microsecond scale.
     47With that said, this is a much fuzzier requirement since it depends on the number of \procs, the number of \ats and the general load of the system.
     48
     49\subsection{Fairness vs Scheduler Locality} \label{fairnessvlocal}
    3650An important performance factor in modern architectures is cache locality. Waiting for data at lower levels or not present in the cache can have a major impact on performance. Having multiple \glspl{hthrd} writing to the same cache lines also leads to cache lines that must be waited on. It is therefore preferable to divide data among each \gls{hthrd}\footnote{This partitioning can be an explicit division up front or using data structures where different \glspl{hthrd} are naturally routed to different cache lines.}.
    3751
    38 For a scheduler, having good locality\footnote{This section discusses \emph{internal locality}, \ie, the locality of the data used by the scheduler versus \emph{external locality}, \ie, how the data used by the application is affected by scheduling. External locality is a much more complicated subject and is discussed in part~\ref{Evaluation} on evaluation.}, \ie, having the data local to each \gls{hthrd}, generally conflicts with fairness. Indeed, good locality often requires avoiding the movement of cache lines, while fairness requires dynamically moving a \gls{thrd}, and as consequence cache lines, to a \gls{hthrd} that is currently available.
     52For a scheduler, having good locality\footnote{This section discusses \emph{internal locality}, \ie, the locality of the data used by the scheduler versus \emph{external locality}, \ie, how the data used by the application is affected by scheduling. External locality is a much more complicated subject and is discussed in the next section.}, \ie, having the data local to each \gls{hthrd}, generally conflicts with fairness. Indeed, good locality often requires avoiding the movement of cache lines, while fairness requires dynamically moving a \gls{thrd}, and as consequence cache lines, to a \gls{hthrd} that is currently available.
    3953
    4054However, I claim that in practice it is possible to strike a balance between fairness and performance because these goals do not necessarily overlap temporally, where Figure~\ref{fig:fair} shows a visual representation of this behaviour. As mentioned, some unfairness is acceptable; therefore it is desirable to have an algorithm that prioritizes cache locality as long as thread delay does not exceed the execution mental-model.
    … …  
    4862\end{figure}
    4963
    50 \section{Design}
     64\subsection{Performance Challenges}\label{pref:challenge}
     65While there exists a multitude of potential scheduling algorithms, they generally always have to contend with the same performance challenges. Since these challenges are recurring themes in the design of a scheduler it is relevant to describe the central ones here before looking at the design.
     66
     67\subsubsection{Scalability}
     68The most basic performance challenge of a scheduler is scalability.
     69Given a large number of \procs and an even larger number of \ats, scalability measures how fast \procs can enqueue and dequeues \ats.
     70One could expect that doubling the number of \procs would double the rate at which \ats are dequeued, but contention on the internal data structure of the scheduler can lead to worst improvements.
     71While the ready-queue itself can be sharded to alleviate the main source of contention, auxillary scheduling features, \eg counting ready \ats, can also be sources of contention.
     72
     73\subsubsection{Migration Cost}
     74Another important source of latency in scheduling is migration.
     75An \at is said to have migrated if it is executed by two different \proc consecutively, which is the process discussed in \ref{fairnessvlocal}.
     76Migrations can have many different causes, but it certain programs it can be all but impossible to limit migrations.
     77Chapter~\ref{microbench} for example, has a benchmark where any \at can potentially unblock any other \at, which can leat to \ats migrating more often than not.
     78Because of this it is important to design the internal data structures of the scheduler to limit the latency penalty from migrations.
     79
     80
     81\section{Inspirations}
    5182In 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 solution to this problem is to shard the ready-queue : create multiple sub-ready-queues that multiple \glspl{hthrd} can access and modify without interfering.
    5283
    53 Before going into the design of \CFA's scheduler proper, I want to discuss two sharding solutions which served as the inspiration scheduler in this thesis.
     84Before going into the design of \CFA's scheduler proper, it is relevant to discuss two sharding solutions which served as the inspiration scheduler in this thesis.
    5485
    5586\subsection{Work-Stealing}
    5687
    57 As I mentioned in \ref{existing:workstealing}, a popular pattern shard the ready-queue is work-stealing. As mentionned, in this pattern each \gls{proc} has its own ready-queue and \glspl{proc} only access each other's ready-queue if they run out of work.
    58 The interesting aspect of workstealing happen in easier scheduling cases, \ie enough work for everyone but no more and no load balancing needed. In these cases, work-stealing is close to optimal scheduling: it can achieve perfect locality and have no contention.
     88As mentioned in \ref{existing:workstealing}, a popular pattern shard the ready-queue is work-stealing.
     89In this pattern each \gls{proc} has its own local ready-queue and \glspl{proc} only access each other's ready-queue if they run out of work on their local ready-queue.
     90The interesting aspect of workstealing happen in easier scheduling cases, \ie enough work for everyone but no more and no load balancing needed.
     91In these cases, work-stealing is close to optimal scheduling: it can achieve perfect locality and have no contention.
    5992On the other hand, work-stealing schedulers only attempt to do load-balancing when a \gls{proc} runs out of work.
    60 This means that the scheduler may never balance unfairness that does not result in a \gls{proc} running out of work.
     93This means that the scheduler never balances unfair loads unless they result in a \gls{proc} running out of work.
    6194Chapter~\ref{microbench} shows that in pathological cases this problem can lead to indefinite starvation.
    6295
    6396
    64 Based on these observation, I conclude that \emph{perfect} scheduler should behave very similarly to work-stealing in the easy cases, but should have more proactive load-balancing if the need arises.
     97Based on these observation, the conclusion is that a \emph{perfect} scheduler should behave very similarly to work-stealing in the easy cases, but should have more proactive load-balancing if the need arises.
    6598
    6699\subsection{Relaxed-Fifo}
    67100An entirely different scheme is to create a ``relaxed-FIFO'' queue as in \todo{cite Trevor's paper}. This approach forgos any ownership between \gls{proc} and ready-queue, and simply creates a pool of ready-queues from which the \glspl{proc} can pick from.
    68101\Glspl{proc} choose ready-queus at random, but timestamps are added to all elements of the queue and dequeues are done by picking two queues and dequeing the oldest element.
     102All subqueues are protected by TryLocks and \procs simply pick a different subqueue if they fail to acquire the TryLock.
    69103The result is a queue that has both decent scalability and sufficient fairness.
    70104The lack of ownership means that as long as one \gls{proc} is still able to repeatedly dequeue elements, it is unlikely that any element will stay on the queue for much longer than any other element.
    … …  
    75109
    76110While the fairness, of this scheme is good, it does suffer in terms of performance.
    77 It requires very wide sharding, \eg at least 4 queues per \gls{hthrd}, and the randomness means locality can suffer significantly and finding non-empty queues can be difficult.
    78 
    79 \section{\CFA}
    80 The \CFA is effectively attempting to merge these two approaches, keeping the best of both.
    81 It is based on the
     111It requires very wide sharding, \eg at least 4 queues per \gls{hthrd}, and finding non-empty queues can be difficult if there are too few ready \ats.
     112
     113\section{Relaxed-FIFO++}
     114Since it has inherent fairness quelities and decent performance in the presence of many \ats, the relaxed-FIFO queue appears as a good candidate to form the basis of a scheduler.
     115The most obvious problems is for workloads where the number of \ats is barely greater than the number of \procs.
     116In these situations, the wide sharding means most of the sub-queues from which the relaxed queue is formed will be empty.
     117The consequence is that when a dequeue operations attempts to pick a sub-queue at random, it is likely that it picks an empty sub-queue and will have to pick again.
     118This problem can repeat an unbounded number of times.
     119
     120As this is the most obvious challenge, it is worth addressing first.
     121The obvious solution is to supplement each subqueue with some sharded data structure that keeps track of which subqueues are empty.
     122This data structure can take many forms, for example simple bitmask or a binary tree that tracks which branch are empty.
     123Following a binary tree on each pick has fairly good Big O complexity and many modern architectures have powerful bitmask manipulation instructions.
     124However, precisely tracking which sub-queues are empty is actually fundamentally problematic.
     125The reason is that each subqueues are already a form of sharding and the sharding width has presumably already chosen to avoid contention.
     126However, tracking which ready queue is empty is only useful if the tracking mechanism uses denser sharding than the sub queues, then it will invariably create a new source of contention.
     127But if the tracking mechanism is not denser than the sub-queues, then it will generally not provide useful because reading this new data structure risks being as costly as simply picking a sub-queue at random.
     128Early experiments with this approach have shown that even with low success rates, randomly picking a sub-queue can be faster than a simple tree walk.
     129
     130The exception to this rule is using local tracking.
     131If each \proc keeps track locally of which sub-queue is empty, then this can be done with a very dense data structure without introducing a new source of contention.
     132The consequence of local tracking however, is that the information is not complete.
     133Each \proc is only aware of the last state it saw each subqueues but does not have any information about freshness.
     134Even on systems with low \gls{hthrd} count, \eg 4 or 8, this can quickly lead to the local information being no better than the random pick.
     135This is due in part to the cost of this maintaining this information and its poor quality.
     136
     137However, using a very low cost approach to local tracking may actually be beneficial.
     138If the local tracking is no more costly than the random pick, than \emph{any} improvement to the succes rate, however low it is, would lead to a performance benefits.
     139This leads to the following approach:
     140
     141\subsection{Dynamic Entropy}\cit{https://xkcd.com/2318/}
     142The Relaxed-FIFO approach can be made to handle the case of mostly empty sub-queues by tweaking the \glsxtrlong{prng}.
     143The \glsxtrshort{prng} state can be seen as containing a list of all the future sub-queues that will be accessed.
     144While this is not particularly useful on its own, the consequence is that if the \glsxtrshort{prng} algorithm can be run \emph{backwards}, then the state also contains a list of all the subqueues that were accessed.
     145Luckily, bidirectional \glsxtrshort{prng} algorithms do exist, for example some Linear Congruential Generators\cit{https://en.wikipedia.org/wiki/Linear\_congruential\_generator} support running the algorithm backwards while offering good quality and performance.
     146This particular \glsxtrshort{prng} can be used as follows:
     147
     148Each \proc maintains two \glsxtrshort{prng} states, which whill be refered to as \texttt{F} and \texttt{B}.
     149
     150When a \proc attempts to dequeue a \at, it picks the subqueues by running the \texttt{B} backwards.
     151When a \proc attempts to enqueue a \at, it runs \texttt{F} forward to pick to subqueue to enqueue to.
     152If the enqueue is successful, the state \texttt{B} is overwritten with the content of \texttt{F}.
     153
     154The result is that each \proc will tend to dequeue \ats that it has itself enqueued.
     155When most sub-queues are empty, this technique increases the odds of finding \ats at very low cost, while also offering an improvement on locality in many cases.
     156
     157However, while this approach does notably improve performance in many cases, this algorithm is still not competitive with work-stealing algorithms.
     158The fundamental problem is that the constant randomness limits how much locality the scheduler offers.
     159This becomes problematic both because the scheduler is likely to get cache misses on internal data-structures and because migration become very frequent.
     160Therefore since the approach of modifying to relaxed-FIFO algorithm to behave more like work stealing does not seem to pan out, the alternative is to do it the other way around.
     161
     162\section{Work Stealing++}
     163To add stronger fairness guarantees to workstealing a few changes.
     164First, the relaxed-FIFO algorithm has fundamentally better fairness because each \proc always monitors all subqueues.
     165Therefore the workstealing algorithm must be prepended with some monitoring.
     166Before attempting to dequeue from a \proc's local queue, the \proc must make some effort to make sure remote queues are not being neglected.
     167To make this possible, \procs must be able to determie which \at has been on the ready-queue the longest.
     168Which is the second aspect that much be added.
     169The relaxed-FIFO approach uses timestamps for each \at and this is also what is done here.
     170
    82171\begin{figure}
    83172        \centering
    84173        \input{base.pstex_t}
    85         \caption[Base \CFA design]{Base \CFA design \smallskip\newline A list of sub-ready queues offers the sharding, two per \glspl{proc}. However, \glspl{proc} can access any of the sub-queues.}
     174        \caption[Base \CFA design]{Base \CFA design \smallskip\newline A Pool of sub-ready queues offers the sharding, two per \glspl{proc}. Each \gls{proc} have local subqueues, however \glspl{proc} can access any of the sub-queues. Each \at is timestamped when enqueued.}
    86175        \label{fig:base}
    87176\end{figure}
    88 
    89 
    90 
    91 % 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.
    92 
    93 % \subsection{Sharding} \label{sec:sharding}
    94 % 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.
    95 
    96 % \begin{figure}
    97 %       \centering
    98 %       \input{base.pstex_t}
    99 %       \caption[Relaxed FIFO list]{Relaxed FIFO list \smallskip\newline List at the base of the scheduler: an array of strictly FIFO lists. The timestamp is in all nodes and cell arrays.}
    100 %       \label{fig:base}
    101 % \end{figure}
    102 
    103 % \subsection{Finding threads}
    104 % Once threads have been distributed onto multiple queues, identifying empty queues becomes a problem. Indeed, if the number of \glspl{thrd} does not far exceed the number of queues, it is probable that several of the cell queues are empty. Figure~\ref{fig:empty} shows an example with 2 \glspl{thrd} running on 8 queues, where the chances of getting an empty queue is 75\% per pick, meaning two random picks yield a \gls{thrd} only half the time. This scenario leads to performance problems since picks that do not yield a \gls{thrd} are not useful and do not necessarily help make more informed guesses.
    105 
    106 % \begin{figure}
    107 %       \centering
    108 %       \input{empty.pstex_t}
    109 %       \caption[``More empty'' Relaxed FIFO list]{``More empty'' Relaxed FIFO list \smallskip\newline Emptier state of the queue: the array contains many empty cells, that is strictly FIFO lists containing no elements.}
    110 %       \label{fig:empty}
    111 % \end{figure}
    112 
    113 % There are several solutions to this problem, but they ultimately all have to encode if a cell has an empty list. My results show the density and locality of this encoding is generally the dominating factor in these scheme. Classic solutions to this problem use one of three techniques to encode the information:
    114 
    115 % \paragraph{Dense Information} Figure~\ref{fig:emptybit} shows a dense bitmask to identify the cell queues currently in use. This approach means processors can often find \glspl{thrd} in constant time, regardless of how many underlying queues are empty. Furthermore, modern x86 CPUs have extended bit manipulation instructions (BMI2) that allow searching the bitmask with very little overhead compared to the randomized selection approach for a filled ready queue, offering good performance even in cases with many empty inner queues. However, this technique has its limits: with a single word\footnote{Word refers here to however many bits can be written atomically.} bitmask, the total amount of ready-queue sharding is limited to the number of bits in the word. With a multi-word bitmask, this maximum limit can be increased arbitrarily, but the look-up time increases. Finally, a dense bitmap, either single or multi-word, causes additional contention problems that reduces performance because of cache misses after updates. This central update bottleneck also means the information in the bitmask is more often stale before a processor can use it to find an item, \ie mask read says there are available \glspl{thrd} but none on queue when the subsequent atomic check is done.
    116 
    117 % \begin{figure}
    118 %       \centering
    119 %       \vspace*{-5pt}
    120 %       {\resizebox{0.75\textwidth}{!}{\input{emptybit.pstex_t}}}
    121 %       \vspace*{-5pt}
    122 %       \caption[Underloaded queue with bitmask]{Underloaded queue with bitmask indicating array cells with items.}
    123 %       \label{fig:emptybit}
    124 
    125 %       \vspace*{10pt}
    126 %       {\resizebox{0.75\textwidth}{!}{\input{emptytree.pstex_t}}}
    127 %       \vspace*{-5pt}
    128 %       \caption[Underloaded queue with binary search-tree]{Underloaded queue with binary search-tree indicating array cells with items.}
    129 %       \label{fig:emptytree}
    130 
    131 %       \vspace*{10pt}
    132 %       {\resizebox{0.95\textwidth}{!}{\input{emptytls.pstex_t}}}
    133 %       \vspace*{-5pt}
    134 %       \caption[Underloaded queue with per processor bitmask]{Underloaded queue with per processor bitmask indicating array cells with items.}
    135 %       \label{fig:emptytls}
    136 % \end{figure}
    137 
    138 % \paragraph{Sparse Information} Figure~\ref{fig:emptytree} shows an approach using a hierarchical tree data-structure to reduce contention and has been shown to work in similar cases~\cite{ellen2007snzi}. However, this approach may lead to poorer performance due to the inherent pointer chasing cost while still allowing significant contention on the nodes of the tree if the tree is shallow.
    139 
    140 % \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.
    141 
    142 % 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.
    143 
    144 % \subsection{Dynamic Entropy}\cit{https://xkcd.com/2318/}
    145 % 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.
    146 
    147 % 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.
    148 
    149 % The algorithm works as follows:
    150 % \begin{itemize}
    151 %       \item Each \gls{proc} has two \glsxtrshort{prng} instances, $F$ and $B$.
    152 %       \item Push and Pop operations occur as discussed in Section~\ref{sec:sharding} with the following exceptions:
    153 %       \begin{itemize}
    154 %               \item Push operations use $F$ going forward on each try and on success $F$ is copied into $B$.
    155 %               \item Pop operations use $B$ going backwards on each try.
    156 %       \end{itemize}
    157 % \end{itemize}
    158 
    159 % The main benefit of this technique is that it basically respects the desired properties of Figure~\ref{fig:fair}. When looking for work, a \gls{proc} first looks at the last cell they pushed to, if any, and then move backwards through its accessed cells. As the \gls{proc} continues looking for work, $F$ moves backwards and $B$ stays in place. As a result, the relation between the two becomes weaker, which means that the probablisitic fairness of the algorithm reverts to normal. Chapter~\ref{proofs} discusses more formally the fairness guarantees of this algorithm.
    160 
    161 % \section{Details}
     177The algorithm is structure as shown in Figure~\ref{fig:base}.
     178This is very similar to classic workstealing except the local queues are placed in an array so \procs can access eachother's queue in constant time.
     179Sharding width can be adjusted based on need.
     180When a \proc attempts to dequeue a \at, it first picks a random remote queue and compares its timestamp to the timestamps of the local queue(s), dequeue from the remote queue if needed.
     181
     182Implemented as as naively state above, this approach has some obvious performance problems.
     183First, it is necessary to have some damping effect on helping.
     184Random effects like cache misses and preemption can add spurious but short bursts of latency for which helping is not helpful, pun intended.
     185The effect of these bursts would be to cause more migrations than needed and make this workstealing approach slowdown to the match the relaxed-FIFO approach.
     186
     187\begin{figure}
     188        \centering
     189        \input{base_avg.pstex_t}
     190        \caption[\CFA design with Moving Average]{\CFA design with Moving Average \smallskip\newline A moving average is added to each subqueue.}
     191        \label{fig:base-ma}
     192\end{figure}
     193
     194A simple solution to this problem is to compare an exponential moving average\cit{https://en.wikipedia.org/wiki/Moving\_average\#Exponential\_moving\_average} instead if the raw timestamps, shown in Figure~\ref{fig:base-ma}.
     195Note that this is slightly more complex than it sounds because since the \at at the head of a subqueue is still waiting, its wait time has not ended.
     196Therefore the exponential moving average is actually an exponential moving average of how long each already dequeued \at have waited.
     197To compare subqueues, the timestamp at the head must be compared to the current time, yielding the bestcase wait time for the \at at the head of the queue.
     198This new waiting is averaged with the stored average.
     199To limit even more the amount of unnecessary migration, a bias can be added to the local queue, where a remote queue is helped only if its moving average is more than \emph{X} times the local queue's average.
     200None of the experimentation that I have run with these scheduler seem to indicate that the choice of the weight for the moving average or the choice of bis is particularly important.
     201Weigths and biases of similar \emph{magnitudes} have similar effects.
     202
     203With these additions to workstealing, scheduling can be made as fair as the relaxed-FIFO approach, well avoiding the majority of unnecessary migrations.
     204Unfortunately, the performance of this approach does suffer in the cases with no risks of starvation.
     205The problem is that the constant polling of remote subqueues generally entail a cache miss.
     206To make things worst, remote subqueues that are very active, \ie \ats are frequently enqueued and dequeued from them, the higher the chances are that polling will incurr a cache-miss.
     207Conversly, the active subqueues do not benefit much from helping since starvation is already a non-issue.
     208This puts this algorithm in an akward situation where it is paying for a cost, but the cost itself suggests the operation was unnecessary.
     209The good news is that this problem can be mitigated
     210
     211\subsection{Redundant Timestamps}
     212The problem with polling remote queues is due to a tension between the consistency requirement on the subqueue.
     213For the subqueues, correctness is critical. There must be a consensus among \procs on which subqueues hold which \ats.
     214Since the timestamps are use for fairness, it is alco important to have consensus and which \at is the oldest.
     215However, when deciding if a remote subqueue is worth polling, correctness is much less of a problem.
     216Since the only need is that a subqueue will eventually be polled, some data staleness can be acceptable.
     217This leads to a tension where stale timestamps are only problematic in some cases.
     218Furthermore, stale timestamps can be somewhat desirable since lower freshness requirements means less tension on the cache coherence protocol.
     219
     220
     221\begin{figure}
     222        \centering
     223        % \input{base_ts2.pstex_t}
     224        \caption[\CFA design with Redundant Timestamps]{\CFA design with Redundant Timestamps \smallskip\newline A array is added containing a copy of the timestamps. These timestamps are written to with relaxed atomics, without fencing, leading to fewer cache invalidations.}
     225        \label{fig:base-ts2}
     226\end{figure}
     227A solution to this is to create a second array containing a copy of the timestamps and average.
     228This copy is updated \emph{after} the subqueue's critical sections using relaxed atomics.
     229\Glspl{proc} now check if polling is needed by comparing the copy of the remote timestamp instead of the actual timestamp.
     230The result is that since there is no fencing, the writes can be buffered and cause fewer cache invalidations.
     231
     232The correctness argument here is somewhat subtle.
     233The data used for deciding whether or not to poll a queue can be stale as long as it does not cause starvation.
     234Therefore, it is acceptable if stale data make queues appear older than they really are but not fresher.
     235For the timestamps, this means that missing writes to the timestamp is acceptable since they will make the head \at look older.
     236For the moving average, as long as the operation are RW-safe, the average is guaranteed to yield a value that is between the oldest and newest values written.
     237Therefore this unprotected read of the timestamp and average satisfy the limited correctness that is required.
     238
     239\begin{figure}
     240        \centering
     241        \input{cache-share.pstex_t}
     242        \caption[CPU design with wide L3 sharing]{CPU design with wide L3 sharing \smallskip\newline A very simple CPU with 4 \glspl{hthrd}. L1 and L2 are private to each \gls{hthrd} but the L3 is shared across to entire core.}
     243        \label{fig:cache-share}
     244\end{figure}
     245
     246\begin{figure}
     247        \centering
     248        \input{cache-noshare.pstex_t}
     249        \caption[CPU design with a narrower L3 sharing]{CPU design with a narrower L3 sharing \smallskip\newline A different CPU design, still with 4 \glspl{hthrd}. L1 and L2 are still private to each \gls{hthrd} but the L3 is shared some of the CPU but there is still two distinct L3 instances.}
     250        \label{fig:cache-noshare}
     251\end{figure}
     252
     253With redundant tiemstamps this scheduling algorithm achieves both the fairness and performance requirements, on some machines.
     254The problem is that the cost of polling and helping is not necessarily consistent across each \gls{hthrd}.
     255For example, on machines where the motherboard holds multiple CPU, cache misses can be satisfied from a cache that belongs to the CPU that missed, the \emph{local} CPU, or by a different CPU, a \emph{remote} one.
     256Cache misses that are satisfied by a remote CPU will have higher latency than if it is satisfied by the local CPU.
     257However, this is not specific to systems with multiple CPUs.
     258Depending on the cache structure, cache-misses can have different latency for the same CPU.
     259The AMD EPYC 7662 CPUs that is described in Chapter~\ref{microbench} is an example of that.
     260Figure~\ref{fig:cache-share} and Figure~\ref{fig:cache-noshare} show two different cache topologies with highlight this difference.
     261In Figure~\ref{fig:cache-share}, all cache instances are either private to a \gls{hthrd} or shared to the entire system, this means latency due to cache-misses are likely fairly consistent.
     262By comparison, in Figure~\ref{fig:cache-noshare} misses in the L2 cache can be satisfied by a hit in either instance of the L3.
     263However, the memory access latency to the remote L3 instance will be notably higher than the memory access latency to the local L3.
     264The impact of these different design on this algorithm is that scheduling will scale very well on architectures similar to Figure~\ref{fig:cache-share}, both will have notably worst scalling with many narrower L3 instances.
     265This is simply because as the number of L3 instances grow, so two does the chances that the random helping will cause significant latency.
     266The solution is to have the scheduler be aware of the cache topology.
     267
     268\subsection{Per CPU Sharding}
     269Building a scheduler that is aware of cache topology poses two main challenges: discovering cache topology and matching \procs to cache instance.
     270Sadly, there is no standard portable way to discover cache topology in C.
     271Therefore, while this is a significant portability challenge, it is outside the scope of this thesis to design a cross-platform cache discovery mechanisms.
     272The rest of this work assumes discovering the cache topology based on Linux's \texttt{/sys/devices/system/cpu} directory.
     273This leaves the challenge of matching \procs to cache instance, or more precisely identifying which subqueues of the ready queue are local to which cache instance.
     274Once this matching is available, the helping algorithm can be changed to add bias so that \procs more often help subqueues local to the same cache instance
     275\footnote{Note that like other biases mentioned in this section, the actual bias value does not appear to need precise tuinng.}.
     276
     277The obvious approach to mapping cache instances to subqueues is to statically tie subqueues to CPUs.
     278Instead of having each subqueue local to a specific \proc, the system is initialized with subqueues for each \glspl{hthrd} up front.
     279Then \procs dequeue and enqueue by first asking which CPU id they are local to, in order to identify which subqueues are the local ones.
     280\Glspl{proc} can get the CPU id from \texttt{sched\_getcpu} or \texttt{librseq}.
     281
     282This approach solves the performance problems on systems with topologies similar to Figure~\ref{fig:cache-noshare}.
     283However, it actually causes some subtle fairness problems in some systems, specifically systems with few \procs and many \glspl{hthrd}.
     284In these cases, the large number of subqueues and the bias agains subqueues tied to different cache instances make it so it is very unlikely any single subqueue is picked.
     285To make things worst, the small number of \procs mean that few helping attempts will be made.
     286This combination of few attempts and low chances make it so a \at stranded on a subqueue that is not actively dequeued from may wait very long before it gets randomly helped.
     287On a system with 2 \procs, 256 \glspl{hthrd} with narrow cache sharing, and a 100:1 bias, it can actually take multiple seconds for a \at to get dequeued from a remote queue.
     288Therefore, a more dynamic matching of subqueues to cache instance is needed.
     289
     290\subsection{Topological Work Stealing}
     291The approach that is used in the \CFA scheduler is to have per-\proc subqueue, but have an excplicit data-structure track which cache instance each subqueue is tied to.
     292This is requires some finess because reading this data structure must lead to fewer cache misses than not having the data structure in the first place.
     293A key element however is that, like the timestamps for helping, reading the cache instance mapping only needs to give the correct result \emph{often enough}.
     294Therefore the algorithm can be built as follows: Before enqueuing or dequeing a \at, each \proc queries the CPU id and the corresponding cache instance.
     295Since subqueues are tied to \procs, each \proc can then update the cache instance mapped to the local subqueue(s).
     296To avoid unnecessary cache line invalidation, the map is only written to if the mapping changes.
     297

  • doc/theses/thierry_delisle_PhD/thesis/text/eval_micro.tex

    rba897d21 r2e9b59b  
    33The first step of evaluation is always to test-out small controlled cases, to ensure that the basics are working properly.
    44This sections presents five different experimental setup, evaluating some of the basic features of \CFA's scheduler.
     5
     6\section{Benchmark Environment}
     7All of these benchmarks are run on two distinct hardware environment, an AMD and an INTEL machine.
     8
     9\paragraph{AMD} The AMD machine is a server with two AMD EPYC 7662 CPUs and 256GB of DDR4 RAM.
     10The server runs Ubuntu 20.04.2 LTS on top of Linux Kernel 5.8.0-55.
     11These EPYCs have 64 cores per CPUs and 2 \glspl{hthrd} per core, for a total of 256 \glspl{hthrd}.
     12The cpus each have 4 MB, 64 MB and 512 MB of L1, L2 and L3 caches respectively.
     13Each L1 and L2 instance are only shared by \glspl{hthrd} on a given core, but each L3 instance is shared by 4 cores, therefore 8 \glspl{hthrd}.
     14
     15\paragraph{Intel} The Intel machine is a server with four Intel Xeon Platinum 8160 CPUs and 384GB of DDR4 RAM.
     16The server runs Ubuntu 20.04.2 LTS on top of Linux Kernel 5.8.0-55.
     17These Xeon Platinums have 24 cores per CPUs and 2 \glspl{hthrd} per core, for a total of 192 \glspl{hthrd}.
     18The cpus each have 3 MB, 96 MB and 132 MB of L1, L2 and L3 caches respectively.
     19Each L1 and L2 instance are only shared by \glspl{hthrd} on a given core, but each L3 instance is shared across the entire CPU, therefore 48 \glspl{hthrd}.
     20
     21This limited sharing of the last level cache on the AMD machine is markedly different than the Intel machine. Indeed, while on both architectures L2 cache misses that are served by L3 caches on a different cpu incurr a significant latency, on AMD it is also the case that cache misses served by a different L3 instance on the same cpu still incur high latency.
     22
    523
    624\section{Cycling latency}
    … …  
    3149\end{figure}
    3250
    33 \todo{check term ``idle sleep handling''}
    3451To avoid this benchmark from being dominated by the idle sleep handling, the number of rings is kept at least as high as the number of \glspl{proc} available.
    3552Beyond this point, adding more rings serves to mitigate even more the idle sleep handling.
    36 This is to avoid the case where one of the worker \glspl{at} runs out of work because of the variation on the number of ready \glspl{at} mentionned above.
     53This is to avoid the case where one of the \glspl{proc} runs out of work because of the variation on the number of ready \glspl{at} mentionned above.
    3754
    3855The actual benchmark is more complicated to handle termination, but that simply requires using a binary semphore or a channel instead of raw \texttt{park}/\texttt{unpark} and carefully picking the order of the \texttt{P} and \texttt{V} with respect to the loop condition.
    3956
    40 \todo{code, setup, results}
    4157\begin{lstlisting}
    4258        Thread.main() {
    … …  
    5268\end{lstlisting}
    5369
     70\begin{figure}
     71        \centering
     72        \input{result.cycle.jax.ops.pstex_t}
     73        \vspace*{-10pt}
     74        \label{fig:cycle:ns:jax}
     75\end{figure}
    5476
    5577\section{Yield}

  • doc/theses/thierry_delisle_PhD/thesis/text/existing.tex

    rba897d21 r2e9b59b  
    22Scheduling is the process of assigning resources to incomming requests.
    33A very common form of this is assigning available workers to work-requests.
    4 The need for scheduling is very common in Computer Science, \eg Operating Systems and Hypervisors schedule available CPUs, NICs schedule available bamdwith, but it is also common in other fields.
    5 For example, assmebly lines are an example of scheduling where parts needed assembly are assigned to line workers.
     4The need for scheduling is very common in Computer Science, \eg Operating Systems and Hypervisors schedule available CPUs, NICs schedule available bamdwith, but scheduling is also common in other fields.
     5For example, in assmebly lines assigning parts in need of assembly to line workers is a form of scheduling.
    66
    77In all these cases, the choice of a scheduling algorithm generally depends first and formost on how much information is available to the scheduler.
    … …  
    1515
    1616\section{Naming Convention}
    17 Scheduling has been studied by various different communities concentrating on different incarnation of the same problems. As a result, their is no real naming convention for scheduling that is respected across these communities. For this document, I will use the term \newterm{task} to refer to the abstract objects being scheduled and the term \newterm{worker} to refer to the objects which will execute these tasks.
     17Scheduling has been studied by various different communities concentrating on different incarnation of the same problems. As a result, their is no real naming convention for scheduling that is respected across these communities. For this document, I will use the term \newterm{\Gls{at}} to refer to the abstract objects being scheduled and the term \newterm{\Gls{proc}} to refer to the objects which will execute these \glspl{at}.
    1818
    1919\section{Static Scheduling}
    20 Static schedulers require that tasks have their dependencies and costs explicitly and exhaustively specified prior schedule.
     20Static schedulers require that \glspl{at} have their dependencies and costs explicitly and exhaustively specified prior schedule.
    2121The scheduler then processes this input ahead of time and producess a \newterm{schedule} to which the system can later adhere.
    2222This approach is generally popular in real-time systems since the need for strong guarantees justifies the cost of supplying this information.
    … …  
    2626
    2727\section{Dynamic Scheduling}
    28 It may be difficult to fulfill the requirements of static scheduler if dependencies are conditionnal. In this case, it may be preferable to detect dependencies at runtime. This detection effectively takes the form of halting or suspending a task with unfulfilled dependencies and adding one or more new task(s) to the system. The new task(s) have the responsability of adding the dependent task back in the system once completed. As a consequence, the scheduler may have an incomplete view of the system, seeing only tasks we no pending dependencies. Schedulers that support this detection at runtime are referred to as \newterm{Dynamic Schedulers}.
     28It may be difficult to fulfill the requirements of static scheduler if dependencies are conditionnal. In this case, it may be preferable to detect dependencies at runtime. This detection effectively takes the form of adding one or more new \gls{at}(s) to the system as their dependencies are resolved. As well as potentially halting or suspending a \gls{at} that dynamically detect unfulfilled dependencies. Each \gls{at} has the responsability of adding the dependent \glspl{at} back in the system once completed. As a consequence, the scheduler may have an incomplete view of the system, seeing only \glspl{at} we no pending dependencies. Schedulers that support this detection at runtime are referred to as \newterm{Dynamic Schedulers}.
    2929
    3030\subsection{Explicitly Informed Dynamic Schedulers}
    31 While dynamic schedulers do not have access to an exhaustive list of dependencies for a task, they may require to provide more or less information about each task, including for example: expected duration, required ressources, relative importance, etc. The scheduler can then use this information to direct the scheduling decisions. \cit{Examples of schedulers with more information} Precisely providing this information can be difficult for programmers, especially \emph{predicted} behaviour, and the scheduler may need to support some amount of imprecision in the provided information. For example, specifying that a tasks takes approximately 5 seconds to complete, rather than exactly 5 seconds. User provided information can also become a significant burden depending how the effort to provide the information scales with the number of tasks and there complexity. For example, providing an exhaustive list of files read by 5 tasks is an easier requirement the providing an exhaustive list of memory addresses accessed by 10'000 distinct tasks.
     31While dynamic schedulers do not have access to an exhaustive list of dependencies for a \gls{at}, they may require to provide more or less information about each \gls{at}, including for example: expected duration, required ressources, relative importance, etc. The scheduler can then use this information to direct the scheduling decisions. \cit{Examples of schedulers with more information} Precisely providing this information can be difficult for programmers, especially \emph{predicted} behaviour, and the scheduler may need to support some amount of imprecision in the provided information. For example, specifying that a \glspl{at} takes approximately 5 seconds to complete, rather than exactly 5 seconds. User provided information can also become a significant burden depending how the effort to provide the information scales with the number of \glspl{at} and there complexity. For example, providing an exhaustive list of files read by 5 \glspl{at} is an easier requirement the providing an exhaustive list of memory addresses accessed by 10'000 distinct \glspl{at}.
    3232
    3333Since the goal of this thesis is to provide a scheduler as a replacement for \CFA's existing \emph{uninformed} scheduler, Explicitly Informed schedulers are less relevant to this project. Nevertheless, some strategies are worth mentionnding.
    3434
    3535\subsubsection{Prority Scheduling}
    36 A commonly used information that schedulers used to direct the algorithm is priorities. Each Task is given a priority and higher-priority tasks are preferred to lower-priority ones. The simplest priority scheduling algorithm is to simply require that every task have a distinct pre-established priority and always run the available task with the highest priority. Asking programmers to provide an exhaustive set of unique priorities can be prohibitive when the system has a large number of tasks. It can therefore be diserable for schedulers to support tasks with identical priorities and/or automatically setting and adjusting priorites for tasks.
     36A commonly used information that schedulers used to direct the algorithm is priorities. Each Task is given a priority and higher-priority \glspl{at} are preferred to lower-priority ones. The simplest priority scheduling algorithm is to simply require that every \gls{at} have a distinct pre-established priority and always run the available \gls{at} with the highest priority. Asking programmers to provide an exhaustive set of unique priorities can be prohibitive when the system has a large number of \glspl{at}. It can therefore be diserable for schedulers to support \glspl{at} with identical priorities and/or automatically setting and adjusting priorites for \glspl{at}. The most common operating some variation on priorities with overlaps and dynamic priority adjustments. For example, Microsoft Windows uses a pair of priorities
     37\cit{https://docs.microsoft.com/en-us/windows/win32/procthread/scheduling-priorities,https://docs.microsoft.com/en-us/windows/win32/taskschd/taskschedulerschema-priority-settingstype-element}, one specified by users out of ten possible options and one adjusted by the system.
    3738
    3839\subsection{Uninformed and Self-Informed Dynamic Schedulers}
    39 Several scheduling algorithms do not require programmers to provide additionnal information on each task, and instead make scheduling decisions based solely on internal state and/or information implicitly gathered by the scheduler.
     40Several scheduling algorithms do not require programmers to provide additionnal information on each \gls{at}, and instead make scheduling decisions based solely on internal state and/or information implicitly gathered by the scheduler.
    4041
    4142
    4243\subsubsection{Feedback Scheduling}
    43 As mentionned, Schedulers may also gather information about each tasks to direct their decisions. This design effectively moves the scheduler to some extent into the realm of \newterm{Control Theory}\cite{wiki:controltheory}. This gathering does not generally involve programmers and as such does not increase programmer burden the same way explicitly provided information may. However, some feedback schedulers do offer the option to programmers to offer additionnal information on certain tasks, in order to direct scheduling decision. The important distinction being whether or not the scheduler can function without this additionnal information.
     44As mentionned, Schedulers may also gather information about each \glspl{at} to direct their decisions. This design effectively moves the scheduler to some extent into the realm of \newterm{Control Theory}\cite{wiki:controltheory}. This gathering does not generally involve programmers and as such does not increase programmer burden the same way explicitly provided information may. However, some feedback schedulers do offer the option to programmers to offer additionnal information on certain \glspl{at}, in order to direct scheduling decision. The important distinction being whether or not the scheduler can function without this additionnal information.
    4445
    4546
    4647\section{Work Stealing}\label{existing:workstealing}
    47 One of the most popular scheduling algorithm in practice (see~\ref{existing:prod}) is work-stealing. This idea, introduce by \cite{DBLP:conf/fpca/BurtonS81}, effectively has each worker work on its local tasks first, but allows the possibility for other workers to steal local tasks if they run out of tasks. \cite{DBLP:conf/focs/Blumofe94} introduced the more familiar incarnation of this, where each workers has queue of tasks to accomplish and workers without tasks steal tasks from random workers. (The Burton and Sleep algorithm had trees of tasks and stole only among neighbours). Blumofe and Leiserson also prove worst case space and time requirements for well-structured computations.
     48One of the most popular scheduling algorithm in practice (see~\ref{existing:prod}) is work-stealing. This idea, introduce by \cite{DBLP:conf/fpca/BurtonS81}, effectively has each worker work on its local \glspl{at} first, but allows the possibility for other workers to steal local \glspl{at} if they run out of \glspl{at}. \cite{DBLP:conf/focs/Blumofe94} introduced the more familiar incarnation of this, where each workers has queue of \glspl{at} to accomplish and workers without \glspl{at} steal \glspl{at} from random workers. (The Burton and Sleep algorithm had trees of \glspl{at} and stole only among neighbours). Blumofe and Leiserson also prove worst case space and time requirements for well-structured computations.
    4849
    4950Many variations of this algorithm have been proposed over the years\cite{DBLP:journals/ijpp/YangH18}, both optmizations of existing implementations and approaches that account for new metrics.
    … …  
    5152\paragraph{Granularity} A significant portion of early Work Stealing research was concentrating on \newterm{Implicit Parellelism}\cite{wiki:implicitpar}. Since the system was responsible to split the work, granularity is a challenge that cannot be left to the programmers (as opposed to \newterm{Explicit Parellelism}\cite{wiki:explicitpar} where the burden can be left to programmers). In general, fine granularity is better for load balancing and coarse granularity reduces communication overhead. The best performance generally means finding a middle ground between the two. Several methods can be employed, but I believe these are less relevant for threads, which are generally explicit and more coarse grained.
    5253
    53 \paragraph{Task Placement} Since modern computers rely heavily on cache hierarchies\cit{Do I need a citation for this}, migrating tasks from one core to another can be .  \cite{DBLP:journals/tpds/SquillanteL93}
     54\paragraph{Task Placement} Since modern computers rely heavily on cache hierarchies\cit{Do I need a citation for this}, migrating \glspl{at} from one core to another can be .  \cite{DBLP:journals/tpds/SquillanteL93}
    5455
    5556\todo{The survey is not great on this subject}
    … …  
    5859
    5960\subsection{Theoretical Results}
    60 There is also a large body of research on the theoretical aspects of work stealing. These evaluate, for example, the cost of migration\cite{DBLP:conf/sigmetrics/SquillanteN91,DBLP:journals/pe/EagerLZ86}, how affinity affects performance\cite{DBLP:journals/tpds/SquillanteL93,DBLP:journals/mst/AcarBB02,DBLP:journals/ipl/SuksompongLS16} and theoretical models for heterogenous systems\cite{DBLP:journals/jpdc/MirchandaneyTS90,DBLP:journals/mst/BenderR02,DBLP:conf/sigmetrics/GastG10}. \cite{DBLP:journals/jacm/BlellochGM99} examine the space bounds of Work Stealing and \cite{DBLP:journals/siamcomp/BerenbrinkFG03} show that for underloaded systems, the scheduler will complete computations in finite time, \ie is \newterm{stable}. Others show that Work-Stealing is applicable to various scheduling contexts\cite{DBLP:journals/mst/AroraBP01,DBLP:journals/anor/TchiboukdjianGT13,DBLP:conf/isaac/TchiboukdjianGTRB10,DBLP:conf/ppopp/AgrawalLS10,DBLP:conf/spaa/AgrawalFLSSU14}. \cite{DBLP:conf/ipps/ColeR13} also studied how Randomized Work Stealing affects false sharing among tasks.
     61There is also a large body of research on the theoretical aspects of work stealing. These evaluate, for example, the cost of migration\cite{DBLP:conf/sigmetrics/SquillanteN91,DBLP:journals/pe/EagerLZ86}, how affinity affects performance\cite{DBLP:journals/tpds/SquillanteL93,DBLP:journals/mst/AcarBB02,DBLP:journals/ipl/SuksompongLS16} and theoretical models for heterogenous systems\cite{DBLP:journals/jpdc/MirchandaneyTS90,DBLP:journals/mst/BenderR02,DBLP:conf/sigmetrics/GastG10}. \cite{DBLP:journals/jacm/BlellochGM99} examine the space bounds of Work Stealing and \cite{DBLP:journals/siamcomp/BerenbrinkFG03} show that for underloaded systems, the scheduler will complete computations in finite time, \ie is \newterm{stable}. Others show that Work-Stealing is applicable to various scheduling contexts\cite{DBLP:journals/mst/AroraBP01,DBLP:journals/anor/TchiboukdjianGT13,DBLP:conf/isaac/TchiboukdjianGTRB10,DBLP:conf/ppopp/AgrawalLS10,DBLP:conf/spaa/AgrawalFLSSU14}. \cite{DBLP:conf/ipps/ColeR13} also studied how Randomized Work Stealing affects false sharing among \glspl{at}.
    6162
    6263However, as \cite{DBLP:journals/ijpp/YangH18} highlights, it is worth mentionning that this theoretical research has mainly focused on ``fully-strict'' computations, \ie workloads that can be fully represented with a Direct Acyclic Graph. It is unclear how well these distributions represent workloads in real world scenarios.
    6364
    6465\section{Preemption}
    65 One last aspect of scheduling worth mentionning is preemption since many schedulers rely on it for some of their guarantees. Preemption is the idea of interrupting tasks that have been running for too long, effectively injecting suspend points in the applications. There are multiple techniques to achieve this but they all aim to have the effect of guaranteeing that suspend points in a task are never further apart than some fixed duration. While this helps schedulers guarantee that no tasks will unfairly monopolize a worker, preemption can effectively added to any scheduler. Therefore, the only interesting aspect of preemption for the design of scheduling is whether or not to require it.
     66One last aspect of scheduling worth mentionning is preemption since many schedulers rely on it for some of their guarantees. Preemption is the idea of interrupting \glspl{at} that have been running for too long, effectively injecting suspend points in the applications. There are multiple techniques to achieve this but they all aim to have the effect of guaranteeing that suspend points in a \gls{at} are never further apart than some fixed duration. While this helps schedulers guarantee that no \glspl{at} will unfairly monopolize a worker, preemption can effectively added to any scheduler. Therefore, the only interesting aspect of preemption for the design of scheduling is whether or not to require it.
    6667
    6768\section{Schedulers in Production}\label{existing:prod}
    … …  
    6970
    7071\subsection{Operating System Schedulers}
    71 Operating System Schedulers tend to be fairly complex schedulers, they generally support some amount of real-time, aim to balance interactive and non-interactive tasks and support for multiple users sharing hardware without requiring these users to cooperate. Here are more details on a few schedulers used in the common operating systems: Linux, FreeBsd, Microsoft Windows and Apple's OS X. The information is less complete for operating systems behind closed source.
     72Operating System Schedulers tend to be fairly complex schedulers, they generally support some amount of real-time, aim to balance interactive and non-interactive \glspl{at} and support for multiple users sharing hardware without requiring these users to cooperate. Here are more details on a few schedulers used in the common operating systems: Linux, FreeBsd, Microsoft Windows and Apple's OS X. The information is less complete for operating systems behind closed source.
    7273
    7374\paragraph{Linux's CFS}
    74 The default scheduler used by Linux (the Completely Fair Scheduler)\cite{MAN:linux/cfs,MAN:linux/cfs2} is a feedback scheduler based on CPU time. For each processor, it constructs a Red-Black tree of tasks waiting to run, ordering them by amount of CPU time spent. The scheduler schedules the task that has spent the least CPU time. It also supports the concept of \newterm{Nice values}, which are effectively multiplicative factors on the CPU time spent. The ordering of tasks is also impacted by a group based notion of fairness, where tasks belonging to groups having spent less CPU time are preferred to tasks beloning to groups having spent more CPU time. Linux achieves load-balancing by regularly monitoring the system state\cite{MAN:linux/cfs/balancing} and using some heuristic on the load (currently CPU time spent in the last millisecond plus decayed version of the previous time slots\cite{MAN:linux/cfs/pelt}.).
     75The default scheduler used by Linux (the Completely Fair Scheduler)\cite{MAN:linux/cfs,MAN:linux/cfs2} is a feedback scheduler based on CPU time. For each processor, it constructs a Red-Black tree of \glspl{at} waiting to run, ordering them by amount of CPU time spent. The scheduler schedules the \gls{at} that has spent the least CPU time. It also supports the concept of \newterm{Nice values}, which are effectively multiplicative factors on the CPU time spent. The ordering of \glspl{at} is also impacted by a group based notion of fairness, where \glspl{at} belonging to groups having spent less CPU time are preferred to \glspl{at} beloning to groups having spent more CPU time. Linux achieves load-balancing by regularly monitoring the system state\cite{MAN:linux/cfs/balancing} and using some heuristic on the load (currently CPU time spent in the last millisecond plus decayed version of the previous time slots\cite{MAN:linux/cfs/pelt}.).
    7576
    76 \cite{DBLP:conf/eurosys/LoziLFGQF16} shows that Linux's CFS also does work-stealing to balance the workload of each processors, but the paper argues this aspect can be improved significantly. The issues highlighted sem to stem from Linux's need to support fairness across tasks \emph{and} across users\footnote{Enforcing fairness across users means, for example, that given two users: one with a single task and the other with one thousand tasks, the user with a single task does not receive one one thousandth of the CPU time.}, increasing the complexity.
     77\cite{DBLP:conf/eurosys/LoziLFGQF16} shows that Linux's CFS also does work-stealing to balance the workload of each processors, but the paper argues this aspect can be improved significantly. The issues highlighted sem to stem from Linux's need to support fairness across \glspl{at} \emph{and} across users\footnote{Enforcing fairness across users means, for example, that given two users: one with a single \gls{at} and the other with one thousand \glspl{at}, the user with a single \gls{at} does not receive one one thousandth of the CPU time.}, increasing the complexity.
    7778
    78 Linux also offers a FIFO scheduler, a real-time schedulerwhich runs the highest-priority task, and a round-robin scheduler, which is an extension of the fifo-scheduler that adds fixed time slices. \cite{MAN:linux/sched}
     79Linux also offers a FIFO scheduler, a real-time schedulerwhich runs the highest-priority \gls{at}, and a round-robin scheduler, which is an extension of the fifo-scheduler that adds fixed time slices. \cite{MAN:linux/sched}
    7980
    8081\paragraph{FreeBSD}
    … …  
    8283
    8384\paragraph{Windows(OS)}
    84 Microsoft's Operating System's Scheduler\cite{MAN:windows/scheduler} is a feedback scheduler with priorities. It supports 32 levels of priorities, some of which are reserved for real-time and prviliged applications. It schedules tasks based on the highest priorities (lowest number) and how much cpu time each tasks have used. The scheduler may also temporarily adjust priorities after certain effects like the completion of I/O requests.
     85Microsoft's Operating System's Scheduler\cite{MAN:windows/scheduler} is a feedback scheduler with priorities. It supports 32 levels of priorities, some of which are reserved for real-time and prviliged applications. It schedules \glspl{at} based on the highest priorities (lowest number) and how much cpu time each \glspl{at} have used. The scheduler may also temporarily adjust priorities after certain effects like the completion of I/O requests.
    8586
    8687\todo{load balancing}
    … …  
    99100
    100101\subsection{User-Level Schedulers}
    101 By comparison, user level schedulers tend to be simpler, gathering fewer metrics and avoid complex notions of fairness. Part of the simplicity is due to the fact that all tasks have the same user, and therefore cooperation is both feasible and probable.
     102By comparison, user level schedulers tend to be simpler, gathering fewer metrics and avoid complex notions of fairness. Part of the simplicity is due to the fact that all \glspl{at} have the same user, and therefore cooperation is both feasible and probable.
    102103\paragraph{Go}
    103104Go's scheduler uses a Randomized Work Stealing algorithm that has a global runqueue(\emph{GRQ}) and each processor(\emph{P}) has both a fixed-size runqueue(\emph{LRQ}) and a high-priority next ``chair'' holding a single element.\cite{GITHUB:go,YTUBE:go} Preemption is present, but only at function call boundaries.
    … …  
    116117
    117118\paragraph{Intel\textregistered ~Threading Building Blocks}
    118 \newterm{Thread Building Blocks}(TBB) is Intel's task parellelism\cite{wiki:taskparallel} framework. It runs tasks or \newterm{jobs}, schedulable objects that must always run to completion, on a pool of worker threads. TBB's scheduler is a variation of Randomized Work Stealing that also supports higher-priority graph-like dependencies\cite{MAN:tbb/scheduler}. It schedules tasks as follows (where \textit{t} is the last task completed):
     119\newterm{Thread Building Blocks}(TBB) is Intel's task parellelism\cite{wiki:taskparallel} framework. It runs \newterm{jobs}, uninterruptable \glspl{at}, schedulable objects that must always run to completion, on a pool of worker threads. TBB's scheduler is a variation of Randomized Work Stealing that also supports higher-priority graph-like dependencies\cite{MAN:tbb/scheduler}. It schedules \glspl{at} as follows (where \textit{t} is the last \gls{at} completed):
    119120\begin{displayquote}
    120121        \begin{enumerate}
    … …  
    136137
    137138\paragraph{Grand Central Dispatch}
    138 This is an API produce by Apple\cit{Official GCD source} that offers task parellelism\cite{wiki:taskparallel}. Its distinctive aspect is that it uses multiple ``Dispatch Queues'', some of which are created by programmers. These queues each have their own local ordering guarantees, \eg tasks on queue $A$ are executed in \emph{FIFO} order.
     139This is an API produce by Apple\cit{Official GCD source} that offers task parellelism\cite{wiki:taskparallel}. Its distinctive aspect is that it uses multiple ``Dispatch Queues'', some of which are created by programmers. These queues each have their own local ordering guarantees, \eg \glspl{at} on queue $A$ are executed in \emph{FIFO} order.
    139140
    140141\todo{load balancing and scheduling}

  • doc/theses/thierry_delisle_PhD/thesis/text/io.tex

    rba897d21 r2e9b59b  
    173173The consequence is that the amount of parallelism used to prepare submissions for the next system call is limited.
    174174Beyond this limit, the length of the system call is the throughput limiting factor.
    175 I concluded from early experiments that preparing submissions seems to take about as long as the system call itself, which means that with a single @io_uring@ instance, there is no benefit in terms of \io throughput to having more than two \glspl{hthrd}.
     175I concluded from early experiments that preparing submissions seems to take at most as long as the system call itself, which means that with a single @io_uring@ instance, there is no benefit in terms of \io throughput to having more than two \glspl{hthrd}.
    176176Therefore the design of the submission engine must manage multiple instances of @io_uring@ running in parallel, effectively sharding @io_uring@ instances.
    177177Similarly to scheduling, this sharding can be done privately, \ie, one instance per \glspl{proc}, in decoupled pools, \ie, a pool of \glspl{proc} use a pool of @io_uring@ instances without one-to-one coupling between any given instance and any given \gls{proc}, or some mix of the two.
    … …  
    200200The only added complexity is that the number of SQEs is fixed, which means allocation can fail.
    201201
    202 Allocation failures need to be pushed up to the routing algorithm: \glspl{thrd} attempting \io operations must not be directed to @io_uring@ instances without sufficient SQEs available.
     202Allocation failures need to be pushed up to a routing algorithm: \glspl{thrd} attempting \io operations must not be directed to @io_uring@ instances without sufficient SQEs available.
    203203Furthermore, the routing algorithm should block operations up-front if none of the instances have available SQEs.
    204204
    … …  
    214214
    215215In the case of designating a \gls{thrd}, ideally, when multiple \glspl{thrd} attempt to submit operations to the same @io_uring@ instance, all requests would be batched together and one of the \glspl{thrd} would do the system call on behalf of the others, referred to as the \newterm{submitter}.
    216 In practice however, it is important that the \io requests are not left pending indefinitely and as such, it may be required to have a current submitter and a next submitter.
     216In practice however, it is important that the \io requests are not left pending indefinitely and as such, it may be required to have a ``next submitter'' that guarentees everything that is missed by the current submitter is seen by the next one.
    217217Indeed, as long as there is a ``next'' submitter, \glspl{thrd} submitting new \io requests can move on, knowing that some future system call will include their request.
    218218Once the system call is done, the submitter must also free SQEs so that the allocator can reused them.
    … …  
    223223If the submission side does not designate submitters, polling can also submit all SQEs as it is polling events.
    224224A simple approach to polling is to allocate a \gls{thrd} per @io_uring@ instance and simply let the poller \glspl{thrd} poll their respective instances when scheduled.
    225 This design is especially convenient for reasons explained in Chapter~\ref{practice}.
    226225
    227226With this pool of instances approach, the big advantage is that it is fairly flexible.
    228227It does not impose restrictions on what \glspl{thrd} submitting \io operations can and cannot do between allocations and submissions.
    229 It also can gracefully handles running out of ressources, SQEs or the kernel returning @EBUSY@.
     228It also can gracefully handle running out of ressources, SQEs or the kernel returning @EBUSY@.
    230229The down side to this is that many of the steps used for submitting need complex synchronization to work properly.
    231230The routing and allocation algorithm needs to keep track of which ring instances have available SQEs, block incoming requests if no instance is available, prevent barging if \glspl{thrd} are already queued up waiting for SQEs and handle SQEs being freed.
    232231The submission side needs to safely append SQEs to the ring buffer, correctly handle chains, make sure no SQE is dropped or left pending forever, notify the allocation side when SQEs can be reused and handle the kernel returning @EBUSY@.
    233 All this synchronization may have a significant cost and, compare to the next approach presented, this synchronization is entirely overhead.
     232All this synchronization may have a significant cost and, compared to the next approach presented, this synchronization is entirely overhead.
    234233
    235234\subsubsection{Private Instances}
    236235Another approach is to simply create one ring instance per \gls{proc}.
    237 This alleviate the need for synchronization on the submissions, requiring only that \glspl{thrd} are not interrupted in between two submission steps.
     236This alleviates the need for synchronization on the submissions, requiring only that \glspl{thrd} are not interrupted in between two submission steps.
    238237This is effectively the same requirement as using @thread_local@ variables.
    239238Since SQEs that are allocated must be submitted to the same ring, on the same \gls{proc}, this effectively forces the application to submit SQEs in allocation order
    … …  
    331330\paragraph{Pending Allocations} can be more complicated to handle.
    332331If the arbiter has available instances, the arbiter can attempt to directly hand over the instance and satisfy the request.
    333 Otherwise
     332Otherwise it must hold onto the list of threads until SQEs are made available again.
     333This handling becomes that much more complex if pending allocation require more than one SQE, since the arbiter must make a decision between statisfying requests in FIFO ordering or satisfy requests for fewer SQEs first.
     334
     335While this arbiter has the potential to solve many of the problems mentionned in above, it also introduces a significant amount of complexity.
     336Tracking which processors are borrowing which instances and which instances have SQEs available ends-up adding a significant synchronization prelude to any I/O operation.
     337Any submission must start with a handshake that pins the currently borrowed instance, if available.
     338An attempt to allocate is then made, but the arbiter can concurrently be attempting to allocate from the same instance from a different \gls{hthrd}.
     339Once the allocation is completed, the submission must still check that the instance is still burrowed before attempt to flush.
     340These extra synchronization steps end-up having a similar cost to the multiple shared instances approach.
     341Furthermore, if the number of instances does not match the number of processors actively submitting I/O, the system can fall into a state where instances are constantly being revoked and end-up cycling the processors, which leads to significant cache deterioration.
     342Because of these reasons, this approach, which sounds promising on paper, does not improve on the private instance approach in practice.
     343
     344\subsubsection{Private Instances V2}
     345
    334346
    335347
    … …  
    394406Finally, the last important part of the \io subsystem is it's interface. There are multiple approaches that can be offered to programmers, each with advantages and disadvantages. The new \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 :
    395407
    396 @ssize_t read(int fd, void *buf, size_t count);@.
     408@ssize_t read(int fd, void *buf, size_t count);@
    397409
    398410\subsection{Replacement}
    399 Replacing the C \glsxtrshort{api}
     411Replacing the C \glsxtrshort{api} is the more intrusive and draconian approach.
     412The goal is to convince the compiler and linker to replace any calls to @read@ to direct them to the \CFA implementation instead of glibc's.
     413This has the advantage of potentially working transparently and supporting existing binaries without needing recompilation.
     414It also offers a, presumably, well known and familiar API that C programmers can simply continue to work with.
     415However, this approach also entails a plethora of subtle technical challenges which generally boils down to making a perfect replacement.
     416If the \CFA interface replaces only \emph{some} of the calls to glibc, then this can easily lead to esoteric concurrency bugs.
     417Since the gcc ecosystems does not offer a scheme for such perfect replacement, this approach was rejected as being laudable but infeasible.
    400418
    401419\subsection{Synchronous Extension}
     420An other interface option is to simply offer an interface that is different in name only. For example:
     421
     422@ssize_t cfa_read(int fd, void *buf, size_t count);@
     423
     424\noindent This is much more feasible but still familiar to C programmers.
     425It comes with the caveat that any code attempting to use it must be recompiled, which can be a big problem considering the amount of existing legacy C binaries.
     426However, it has the advantage of implementation simplicity.
    402427
    403428\subsection{Asynchronous Extension}
     429It is important to mention that there is a certain irony to using only synchronous, therefore blocking, interfaces for a feature often referred to as ``non-blocking'' \io.
     430A fairly traditional way of doing this is using futures\cit{wikipedia futures}.
     431As simple way of doing so is as follows:
     432
     433@future(ssize_t) read(int fd, void *buf, size_t count);@
     434
     435\noindent Note that this approach is not necessarily the most idiomatic usage of futures.
     436The definition of read above ``returns'' the read content through an output parameter which cannot be synchronized on.
     437A more classical asynchronous API could look more like:
     438
     439@future([ssize_t, void *]) read(int fd, size_t count);@
     440
     441\noindent However, this interface immediately introduces memory lifetime challenges since the call must effectively allocate a buffer to be returned.
     442Because of the performance implications of this, the first approach is considered preferable as it is more familiar to C programmers.
    404443
    405444\subsection{Interface directly to \lstinline{io_uring}}
     445Finally, an other interface that can be relevant is to simply expose directly the underlying \texttt{io\_uring} interface. For example:
     446
     447@array(SQE, want) cfa_io_allocate(int want);@
     448
     449@void cfa_io_submit( const array(SQE, have) & );@
     450
     451\noindent This offers more flexibility to users wanting to fully use all of the \texttt{io\_uring} features.
     452However, it is not the most user-friendly option.
     453It obviously imposes a strong dependency between user code and \texttt{io\_uring} but at the same time restricting users to usages that are compatible with how \CFA internally uses \texttt{io\_uring}.
     454
     455

  • doc/theses/thierry_delisle_PhD/thesis/text/practice.tex

    rba897d21 r2e9b59b  
    22The scheduling algorithm discribed in Chapter~\ref{core} addresses scheduling in a stable state.
    33However, it does not address problems that occur when the system changes state.
    4 Indeed the \CFA runtime, supports expanding and shrinking the number of KTHREAD\_place \todo{add kthrd to glossary}, both manually and, to some extent automatically.
     4Indeed the \CFA runtime, supports expanding and shrinking the number of \procs, both manually and, to some extent, automatically.
    55This entails that the scheduling algorithm must support these transitions.
    66
    7 \section{Resizing}
     7More precise \CFA supports adding \procs using the RAII object @processor@.
     8These objects can be created at any time and can be destroyed at any time.
     9They are normally create as automatic stack variables, but this is not a requirement.
     10
     11The consequence is that the scheduler and \io subsystems must support \procs comming in and out of existence.
     12
     13\section{Manual Resizing}
     14The consequence of dynamically changing the number of \procs is that all internal arrays that are sized based on the number of \procs neede to be \texttt{realloc}ed.
     15This also means that any references into these arrays, pointers or indexes, may need to be fixed when shrinking\footnote{Indexes may still need fixing because there is no guarantee the \proc causing the shrink had the highest index. Therefore indexes need to be reassigned to preserve contiguous indexes.}.
     16
     17There are no performance requirements, within reason, for resizing since this is usually considered as part of setup and teardown.
     18However, this operation has strict correctness requirements since shrinking and idle sleep can easily lead to deadlocks.
     19It should also avoid as much as possible any effect on performance when the number of \procs remain constant.
     20This later requirement prehibits simple solutions, like simply adding a global lock to these arrays.
     21
     22\subsection{Read-Copy-Update}
     23One solution is to use the Read-Copy-Update\cite{wiki:rcu} pattern.
     24In this pattern, resizing is done by creating a copy of the internal data strucures, updating the copy with the desired changes, and then attempt an Idiana Jones Switch to replace the original witht the copy.
     25This approach potentially has the advantage that it may not need any synchronization to do the switch.
     26The switch definitely implies a race where \procs could still use the previous, original, data structure after the copy was switched in.
     27The important question then becomes whether or not this race can be recovered from.
     28If the changes that arrived late can be transferred from the original to the copy then this solution works.
     29
     30For linked-lists, dequeing is somewhat of a problem.
     31Dequeing from the original will not necessarily update the copy which could lead to multiple \procs dequeing the same \at.
     32Fixing this requires making the array contain pointers to subqueues rather than the subqueues themselves.
     33
     34Another challenge is that the original must be kept until all \procs have witnessed the change.
     35This is a straight forward memory reclamation challenge but it does mean that every operation will need \emph{some} form of synchronization.
     36If each of these operation does need synchronization then it is possible a simpler solution achieves the same performance.
     37Because in addition to the classic challenge of memory reclamation, transferring the original data to the copy before reclaiming it poses additional challenges.
     38Especially merging subqueues while having a minimal impact on fairness and locality.
     39
     40\subsection{Read-Writer Lock}
     41A simpler approach would be to use a \newterm{Readers-Writer Lock}\cite{wiki:rwlock} where the resizing requires acquiring the lock as a writer while simply enqueing/dequeing \ats requires acquiring the lock as a reader.
     42Using a Readers-Writer lock solves the problem of dynamically resizing and leaves the challenge of finding or building a lock with sufficient good read-side performance.
     43Since this is not a very complex challenge and an ad-hoc solution is perfectly acceptable, building a Readers-Writer lock was the path taken.
     44
     45To maximize reader scalability, the readers should not contend with eachother when attempting to acquire and release the critical sections.
     46This effectively requires that each reader have its own piece of memory to mark as locked and unlocked.
     47Reades then acquire the lock wait for writers to finish the critical section and then acquire their local spinlocks.
     48Writers acquire the global lock, so writers have mutual exclusion among themselves, and then acquires each of the local reader locks.
     49Acquiring all the local locks guarantees mutual exclusion between the readers and the writer, while the wait on the read side prevents readers from continously starving the writer.
     50\todo{reference listings}
     51
     52\begin{lstlisting}
     53void read_lock() {
     54        // Step 1 : make sure no writers in
     55        while write_lock { Pause(); }
     56
     57        // May need fence here
     58
     59        // Step 2 : acquire our local lock
     60        while atomic_xchg( tls.lock ) {
     61                Pause();
     62        }
     63}
     64
     65void read_unlock() {
     66        tls.lock = false;
     67}
     68\end{lstlisting}
     69
     70\begin{lstlisting}
     71void write_lock()  {
     72        // Step 1 : lock global lock
     73        while atomic_xchg( write_lock ) {
     74                Pause();
     75        }
     76
     77        // Step 2 : lock per-proc locks
     78        for t in all_tls {
     79                while atomic_xchg( t.lock ) {
     80                        Pause();
     81                }
     82        }
     83}
     84
     85void write_unlock() {
     86        // Step 1 : release local locks
     87        for t in all_tls {
     88                t.lock = false;
     89        }
     90
     91        // Step 2 : release global lock
     92        write_lock = false;
     93}
     94\end{lstlisting}
    895
    996\section{Idle-Sleep}
     97In addition to users manually changing the number of \procs, it is desireable to support ``removing'' \procs when there is not enough \ats for all the \procs to be useful.
     98While manual resizing is expected to be rare, the number of \ats is expected to vary much more which means \procs may need to be ``removed'' for only short periods of time.
     99Furthermore, race conditions that spuriously lead to the impression no \ats are ready are actually common in practice.
     100Therefore \procs should not be actually \emph{removed} but simply put into an idle state where the \gls{kthrd} is blocked until more \ats become ready.
     101This state is referred to as \newterm{Idle-Sleep}.
     102
     103Idle sleep effectively encompasses several challenges.
     104First some data structure needs to keep track of all \procs that are in idle sleep.
     105Because of idle sleep can be spurious, this data structure has strict performance requirements in addition to the strict correctness requirements.
     106Next, some tool must be used to block kernel threads \glspl{kthrd}, \eg \texttt{pthread\_cond\_wait}, pthread semaphores.
     107The complexity here is to support \at parking and unparking, timers, \io operations and all other \CFA features with minimal complexity.
     108Finally, idle sleep also includes a heuristic to determine the appropriate number of \procs to be in idle sleep an any given time.
     109This third challenge is however outside the scope of this thesis because developping a general heuristic is involved enough to justify its own work.
     110The \CFA scheduler simply follows the ``Race-to-Idle'\cit{https://doi.org/10.1137/1.9781611973099.100}' approach where a sleeping \proc is woken any time an \at becomes ready and \procs go to idle sleep anytime they run out of work.
     111
     112
     113\section{Tracking Sleepers}
     114Tracking which \procs are in idle sleep requires a data structure holding all the sleeping \procs, but more importantly it requires a concurrent \emph{handshake} so that no \at is stranded on a ready-queue with no active \proc.
     115The classic challenge is when a \at is made ready while a \proc is going to sleep, there is a race where the new \at may not see the sleeping \proc and the sleeping \proc may not see the ready \at.
     116
     117Furthermore, the ``Race-to-Idle'' approach means that there is some
     118
     119\section{Sleeping}
     120
     121\subsection{Event FDs}
     122
     123\subsection{Epoll}
     124
     125\subsection{\texttt{io\_uring}}
     126
     127\section{Reducing Latency}

  • doc/theses/thierry_delisle_PhD/thesis/thesis.tex

    rba897d21 r2e9b59b  
    202202
    203203\newcommand\io{\glsxtrshort{io}\xspace}%
     204\newcommand\at{\gls{at}\xspace}%
     205\newcommand\ats{\glspl{at}\xspace}%
     206\newcommand\proc{\gls{proc}\xspace}%
     207\newcommand\procs{\glspl{proc}\xspace}%
    204208
    205209%======================================================================
Note: See TracChangeset for help on using the changeset viewer.