Changes in / [683cc13:36cc24a]
- Location:
- doc/theses/thierry_delisle_PhD/thesis/text
- Files:
-
- 4 edited
Legend:
- Unmodified
- Added
- Removed
-
doc/theses/thierry_delisle_PhD/thesis/text/eval_macro.tex
r683cc13 r36cc24a 33 33 This models adds flexibility to the implementation, as the serving logic can now block on user-level primitives without affecting other connections. 34 34 35 Memcached is not built according to a thread-per-connection model, but there exists a port of it that is, which was built for libfibre in\cite{DBLP:journals/pomacs/KarstenB20}.35 Memcached is not built according to a thread-per-connection model, but there exists a port of it that is, which was built for libfibre~\cite{DBLP:journals/pomacs/KarstenB20}. 36 36 Therefore this version can both be compared to the original version and to a port to the \CFA runtime. 37 37 … … 39 39 \begin{itemize} 40 40 \item \emph{vanilla}: the official release of memcached, version~1.6.9. 41 \item \emph{fibre}: a modification of vanilla which uses the thread per connection model on top of the libfibre runtime ~\cite{DBLP:journals/pomacs/KarstenB20}.41 \item \emph{fibre}: a modification of vanilla which uses the thread per connection model on top of the libfibre runtime. 42 42 \item \emph{cfa}: a modification of the fibre webserver that replaces the libfibre runtime with \CFA. 43 43 \end{itemize} -
doc/theses/thierry_delisle_PhD/thesis/text/eval_micro.tex
r683cc13 r36cc24a 1 1 \chapter{Micro-Benchmarks}\label{microbench} 2 2 3 The first step in evaluating this work is to test-out small controlled cases to ensure the basics work properly. 4 This chapter presents five different experimental setup, evaluating some of the basic features of \CFA's scheduler. 3 The first step in evaluating this work is to test small controlled cases to ensure the basics work properly. 4 This chapter presents five different experimental setups, evaluating the basic features of the \CFA, libfibre~\cite{libfibre}, Go, and Tokio~\cite{Tokio} schedulers. 5 All of these systems have a \gls{uthrding} model. 6 Note, all tests in each system are functionally identical and available online~\cite{SchedulingBenchmarks}. 5 7 6 8 \section{Benchmark Environment}\label{microenv} … … 20 22 \end{description} 21 23 22 For all benchmarks, @taskset@ is used to limit the experiment to 1 NUMA Node with no hyper threading.23 If more \glspl{hthrd} are needed, then 1 NUMA Node with hyperthreading is used.24 If still more \glspl{hthrd} are needed, then the experiment is limited to as few NUMA Nodes as needed.24 For all benchmarks, @taskset@ is used to limit the experiment to 1 NUMA node with no hyper threading. 25 If more \glspl{hthrd} are needed, then 1 NUMA node with hyperthreading is used. 26 If still more \glspl{hthrd} are needed, then the experiment is limited to as few NUMA nodes as needed. 25 27 26 28 The limited sharing of the last-level cache on the AMD machine is markedly different than the Intel machine. 27 Indeed, while on both architectures L2 cache misses that are served by L3 caches on a different CPU incur a significant latency, on the AMD it is also the case that cache misses served by a different L3 instance on the same CPU stillincur high latency.29 Indeed, while on both architectures L2 cache misses that are served by L3 caches on a different CPU incur a significant latency, on the AMD it is also the case that cache misses served by a different L3 instance on the same CPU also incur high latency. 28 30 29 31 30 32 \section{Cycling latency} 33 34 The most basic evaluation of any ready queue is the latency needed to push and pop one element from the ready queue. 35 Since these two operations also describe a @yield@ operation, many systems use this operation as the fundamental benchmark. 36 However, yielding can be treated as a special case by optimizing it away since the number of ready \ats does not change. 37 Hence, systems that perform this optimization have an artificial performance benefit because the yield becomes a \emph{nop}. 38 For this reason, I chose a different first benchmark, called \newterm{Cycle Benchmark}. 39 This benchmark arranges a number of \ats into a ring, as seen in Figure~\ref{fig:cycle}, where the ring is a circular singly-linked list. 40 At runtime, each \at unparks the next \at before parking itself. 41 Unparking the next \at pushes that \at onto the ready queue while the ensuing park leads to a \at being popped from the ready queue. 42 31 43 \begin{figure} 32 44 \centering … … 35 47 \label{fig:cycle} 36 48 \end{figure} 37 The most basic evaluation of any ready queue is to evaluate the latency needed to push and pop one element from the ready queue. 38 Since these two operation also describe a @yield@ operation, many systems use this operation as the most basic benchmark. 39 However, yielding can be treated as a special case and some aspects of the scheduler can be optimized away since the number of ready \ats does not change. 40 Not all systems perform this type of optimization, but those that do have an artificial performance benefit because the yield becomes a \emph{nop}. 41 For this reason, I chose a different first benchmark, called \newterm{Cycle Benchmark}. 42 This benchmark arranges a number of \ats into a ring, as seen in Figure~\ref{fig:cycle}, where the ring is a circular singly-linked list. 43 At runtime, each \at unparks the next \at before parking itself. 44 Unparking the next \at pushes that \at onto the ready queue while the ensuing park leads to a \at being popped from the ready queue. 45 46 Hence, the underlying runtime cannot rely on the number of ready \ats staying constant over the duration of the experiment. 47 In fact, the total number of \ats waiting on the ready queue is expected to vary because of the delay between the next \at unparking and the current \at parking. 48 That is, the runtime cannot anticipate that the current task will immediately park. 49 As well, the size of the cycle is also decided based on this delay. 50 Note that, an unpark is like a V on a semaphore, so the subsequent park (P) may not block. 51 If this happens, the scheduler push and pop are avoided and the results of the experiment would be skewed. 52 Because of time-slicing or because cycles can be spread over multiple \procs, a small cycle may see the chain of unparks go full circle before the first \at parks. 53 Every runtime system must handle this race and but cannot optimized away the ready-queue pushes and pops if the cycle is long enough. 49 50 Therefore, the underlying runtime cannot rely on the number of ready \ats staying constant over the duration of the experiment. 51 In fact, the total number of \ats waiting on the ready queue is expected to vary because of the race between the next \at unparking and the current \at parking. 52 That is, the runtime cannot anticipate that the current task immediately parks. 53 As well, the size of the cycle is also decided based on this race, \eg a small cycle may see the chain of unparks go full circle before the first \at parks because of time-slicing or multiple \procs. 54 If this happens, the scheduler push and pop are avoided and the results of the experiment are skewed. 55 (Note, an unpark is like a V on a semaphore, so the subsequent park (P) may not block.) 56 Every runtime system must handle this race and cannot optimized away the ready-queue pushes and pops. 54 57 To prevent any attempt of silently omitting ready-queue operations, the ring of \ats is made big enough so the \ats have time to fully park before being unparked again. 55 58 Finally, to further mitigate any underlying push/pop optimizations, especially on SMP machines, multiple rings are created in the experiment. … … 73 76 \caption[Cycle Benchmark : Pseudo Code]{Cycle Benchmark : Pseudo Code} 74 77 \label{fig:cycle:code} 78 %\end{figure} 79 80 \bigskip 81 82 %\begin{figure} 83 \subfloat[][Throughput, 100 cycles per \proc]{ 84 \resizebox{0.5\linewidth}{!}{ 85 \input{result.cycle.jax.ops.pstex_t} 86 } 87 \label{fig:cycle:jax:ops} 88 } 89 \subfloat[][Throughput, 1 cycle per \proc]{ 90 \resizebox{0.5\linewidth}{!}{ 91 \input{result.cycle.low.jax.ops.pstex_t} 92 } 93 \label{fig:cycle:jax:low:ops} 94 } 95 96 \subfloat[][Scalability, 100 cycles per \proc]{ 97 \resizebox{0.5\linewidth}{!}{ 98 \input{result.cycle.jax.ns.pstex_t} 99 } 100 \label{fig:cycle:jax:ns} 101 } 102 \subfloat[][Scalability, 1 cycle per \proc]{ 103 \resizebox{0.5\linewidth}{!}{ 104 \input{result.cycle.low.jax.ns.pstex_t} 105 } 106 \label{fig:cycle:jax:low:ns} 107 } 108 \caption[Cycle Benchmark on Intel]{Cycle Benchmark on Intel\smallskip\newline Throughput and scalability as a function of \proc count, 5 \ats per cycle, and different cycle count. For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are maximums while the solid line is the medium.} 109 \label{fig:cycle:jax} 75 110 \end{figure} 76 111 77 112 \subsection{Results} 78 \begin{figure} 79 \subfloat[][Throughput, 100 cycles per \proc]{ 80 \resizebox{0.5\linewidth}{!}{ 81 \input{result.cycle.jax.ops.pstex_t} 82 } 83 \label{fig:cycle:jax:ops} 84 } 85 \subfloat[][Throughput, 1 cycle per \proc]{ 86 \resizebox{0.5\linewidth}{!}{ 87 \input{result.cycle.low.jax.ops.pstex_t} 88 } 89 \label{fig:cycle:jax:low:ops} 90 } 91 92 \subfloat[][Scalability, 100 cycles per \proc]{ 93 \resizebox{0.5\linewidth}{!}{ 94 \input{result.cycle.jax.ns.pstex_t} 95 } 96 \label{fig:cycle:jax:ns} 97 } 98 \subfloat[][Scalability, 1 cycle per \proc]{ 99 \resizebox{0.5\linewidth}{!}{ 100 \input{result.cycle.low.jax.ns.pstex_t} 101 } 102 \label{fig:cycle:jax:low:ns} 103 } 104 \caption[Cycle Benchmark on Intel]{Cycle Benchmark on Intel\smallskip\newline Throughput and Scalability as a function of \proc count 5 \ats per cycle and different cycle count. For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.} 105 \label{fig:cycle:jax} 106 \end{figure} 107 108 \begin{figure} 109 \subfloat[][Throughput, 100 cycles per \proc]{ 110 \resizebox{0.5\linewidth}{!}{ 111 \input{result.cycle.nasus.ops.pstex_t} 112 } 113 \label{fig:cycle:nasus:ops} 114 } 115 \subfloat[][Throughput, 1 cycle per \proc]{ 116 \resizebox{0.5\linewidth}{!}{ 117 \input{result.cycle.low.nasus.ops.pstex_t} 118 } 119 \label{fig:cycle:nasus:low:ops} 120 } 121 122 \subfloat[][Scalability, 100 cycles per \proc]{ 123 \resizebox{0.5\linewidth}{!}{ 124 \input{result.cycle.nasus.ns.pstex_t} 125 } 126 \label{fig:cycle:nasus:ns} 127 } 128 \subfloat[][Scalability, 1 cycle per \proc]{ 129 \resizebox{0.5\linewidth}{!}{ 130 \input{result.cycle.low.nasus.ns.pstex_t} 131 } 132 \label{fig:cycle:nasus:low:ns} 133 } 134 \caption[Cycle Benchmark on AMD]{Cycle Benchmark on AMD\smallskip\newline Throughput and Scalability as a function of \proc count 5 \ats per cycle and different cycle count. For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.} 135 \label{fig:cycle:nasus} 136 \end{figure} 137 Figure~\ref{fig:cycle:jax} and Figure~\ref{fig:cycle:nasus} shows the throughput as a function of \proc count on Intel and AMD respectively, where each cycle has 5 \ats. 113 114 Figures~\ref{fig:cycle:jax} and~\ref{fig:cycle:nasus} show the throughput as a function of \proc count on Intel and AMD respectively, where each cycle has 5 \ats. 138 115 The graphs show traditional throughput on the top row and \newterm{scalability} on the bottom row. 139 Where scalability uses the same databut the Y axis is calculated as the number of \procs over the throughput.116 Scalability uses the same data as throughput but the Y axis is calculated as the number of \procs over the throughput. 140 117 In this representation, perfect scalability should appear as a horizontal line, \eg, if doubling the number of \procs doubles the throughput, then the relation stays the same. 141 118 The left column shows results for 100 cycles per \proc, enough cycles to always keep every \proc busy. 142 The right column shows results for only 1 cycle per \proc, where the ready queues are expected to be near empty at all times. 143 The distinction is meaningful because the idle sleep subsystem is expected to matter only in the right column, where spurious effects can cause a \proc to run out of work temporarily. 144 145 The experiment was run 15 times for each series and processor count and the markers on the graph show all of the results obtained. 146 Each series also has a solid and two dashed lines highlighting the median, maximum and minimum result respectively. 147 This presentation offers an overview of the distribution of the results for each series. 119 The right column shows results for 1 cycle per \proc, where the ready queues are expected to be near empty most of the time. 120 The distinction between 100 and 1 cycles is meaningful because the idle sleep subsystem is expected to matter only in the right column, where spurious effects can cause a \proc to run out of work temporarily. 121 122 \begin{figure} 123 \subfloat[][Throughput, 100 cycles per \proc]{ 124 \resizebox{0.5\linewidth}{!}{ 125 \input{result.cycle.nasus.ops.pstex_t} 126 } 127 \label{fig:cycle:nasus:ops} 128 } 129 \subfloat[][Throughput, 1 cycle per \proc]{ 130 \resizebox{0.5\linewidth}{!}{ 131 \input{result.cycle.low.nasus.ops.pstex_t} 132 } 133 \label{fig:cycle:nasus:low:ops} 134 } 135 136 \subfloat[][Scalability, 100 cycles per \proc]{ 137 \resizebox{0.5\linewidth}{!}{ 138 \input{result.cycle.nasus.ns.pstex_t} 139 } 140 \label{fig:cycle:nasus:ns} 141 } 142 \subfloat[][Scalability, 1 cycle per \proc]{ 143 \resizebox{0.5\linewidth}{!}{ 144 \input{result.cycle.low.nasus.ns.pstex_t} 145 } 146 \label{fig:cycle:nasus:low:ns} 147 } 148 \caption[Cycle Benchmark on AMD]{Cycle Benchmark on AMD\smallskip\newline Throughput and scalability as a function of \proc count, 5 \ats per cycle, and different cycle count. For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are extremes while the solid line is the medium.} 149 \label{fig:cycle:nasus} 150 \end{figure} 151 152 The experiment ran 15 times for each series and processor count. 153 Each series has a solid and two dashed lines representing the median, maximum and minimum result respectively, where the minimum/maximum lines are referred to as the \emph{extremes}.\footnote{ 154 An alternative display is to use error bars with min/max as the bottom/top for the bar. 155 However, this approach is not truly an error bar around a mean value and I felt the connected lines are easier to read.} 156 This graph presentation offers an overview of the distribution of the results for each series. 148 157 149 158 The experimental setup uses taskset to limit the placement of \glspl{kthrd} by the operating system. 150 As mentioned in Section~\ref{microenv}, the experi ement is setup to prioritize running on 2\glspl{hthrd} per core before running on multiple sockets.151 For the Intel machine, this means that from 1 to 24 \procs , one socket and \emph{no} hyperthreading is used and from 25 to 48 \procs, still only one socketbut \emph{with} hyperthreading.159 As mentioned in Section~\ref{microenv}, the experiment is setup to prioritize running on two \glspl{hthrd} per core before running on multiple sockets. 160 For the Intel machine, this means that from 1 to 24 \procs one socket and \emph{no} hyperthreading is used, and from 25 to 48 \procs still only one socket is used but \emph{with} hyperthreading. 152 161 This pattern is repeated between 49 and 96, between 97 and 144, and between 145 and 192. 153 162 On AMD, the same algorithm is used, but the machine only has 2 sockets. 154 So hyperthreading\footnote{Hyperthreading normally refers specifically to the technique used by Intel, however here it is loosely used to refer to AMD's equivalent feature.} is used when the \proc count reach 65 and 193. 155 163 So hyperthreading\footnote{ 164 Hyperthreading normally refers specifically to the technique used by Intel, however it is often used generically to refer to any equivalent feature.} 165 is used when the \proc count reach 65 and 193. 166 167 The performance goal of \CFA is to obtain equivalent performance to other less fair schedulers. 156 168 Figure~\ref{fig:cycle:jax:ops} and Figure~\ref{fig:cycle:jax:ns} show that for 100 cycles per \proc, \CFA, Go and Tokio all obtain effectively the same performance. 157 169 Libfibre is slightly behind in this case but still scales decently. 158 As a result of the \gls{kthrd} placement, we can see thatadditional \procs from 25 to 48 offer less performance improvements for all runtimes.170 As a result of the \gls{kthrd} placement, additional \procs from 25 to 48 offer less performance improvements for all runtimes. 159 171 As expected, this pattern repeats between \proc count 72 and 96. 160 The performance goal of \CFA is to obtain equivalent performance to other, less fair schedulers and that is what results show. 161 Figure~\ref{fig:cycle:jax:ops} and \ref{fig:cycle:jax:ns} show very good throughput and scalability for all runtimes. 172 Hence, Figures~\ref{fig:cycle:jax:ops} and \ref{fig:cycle:jax:ns} show very good throughput and scalability for all runtimes. 162 173 163 174 When running only a single cycle, the story is slightly different. 164 \CFA and tokio obtain very smiliar results overall, but tokio shows notably more variations in the results.165 While \CFA, Go and tokio achive equivalent performance with 100 cycles per \proc, with only 1 cycle per \procGo achieves slightly better performance.175 \CFA and Tokio obtain very similar results overall, but Tokio shows notably more variations in the results. 176 While \CFA, Go and Tokio achieve equivalent performance with 100 cycles per \proc, with only 1 cycle per \proc, Go achieves slightly better performance. 166 177 This difference in throughput and scalability is due to the idle-sleep mechanism. 167 With very few cycles, stealing or helping can cause a cascade of tasks migration and trick \proc into very short idle sleeps. 168 Both effect will negatively affect performance. 169 170 An interesting and unusual result is that libfibre achieves better performance with fewer cycle. 178 With very few cycles, stealing or helping can cause a cascade of tasks migration and trick a \proc into very short idle sleeps, which negatively affect performance. 179 180 An interesting and unusual result is that libfibre achieves better performance with 1 cycle. 171 181 This suggest that the cascade effect is never present in libfibre and that some bottleneck disappears in this context. 172 182 However, I did not investigate this result any deeper. 173 183 174 184 Figure~\ref{fig:cycle:nasus} show a similar story happening on AMD as it does on Intel. 175 The different performance improvements and plateaus due to cache topologyappear at the expected \proc counts of 64, 128 and 192, for the same reasons as on Intel.176 Unlike Intel, on AMD all 4 runtimes achieve very similar throughput and scalability for 100 cycles per \proc .185 The different performance improvements and plateaus are due to cache topology and appear at the expected \proc counts of 64, 128 and 192, for the same reasons as on Intel. 186 Unlike Intel, on AMD all 4 runtimes achieve very similar throughput and scalability for 100 cycles per \proc, with some variations in the results. 177 187 178 188 In the 1 cycle per \proc experiment, the same performance increase for libfibre is visible. 179 However, unlike on Intel, tokio achieves the same performance as Go rather than \CFA.189 However, unlike on Intel, Tokio achieves the same performance as Go rather than \CFA. 180 190 This leaves \CFA trailing behind in this particular case, but only at hight core counts. 181 Presumably this is because in this case, \emph{any} helping is likely to cause a cascade of \procs running out of work and attempting to steal. 182 Since this effect is only problematic in cases with 1 \at per \proc it is not very meaningful for the general performance. 183 184 The conclusion from both architectures is that all of the compared runtime have fairly equivalent performance in this scenario. 185 Which demonstrate that in this case \CFA achieves equivalent performance. 191 For this case, \emph{any} helping is likely to cause a cascade of \procs running out of work and attempting to steal. 192 Essentially, \CFA's fairness must result in slower performance in some workload. 193 Fortunately, this effect is only problematic in pathological cases, \eg with 1 \at per \proc, which seldom occurs in most workloads. 194 195 The conclusion from both architectures is that all of the compared runtime have fairly equivalent performance for this micro-benchmark. 196 This result shows that the \CFA scheduler has achieved the goal of obtaining equivalent performance to other less fair schedulers. 186 197 187 198 \section{Yield} 188 199 For completion, the classic yield benchmark is included. 189 This benchmark is simpler than the cycle test: it creates many \ats that call @yield@. 200 Here, the throughput is dominated by the mechanism used to handle the @yield@ function. 201 Figure~\ref{fig:yield:code} shows pseudo code for this benchmark, where the cycle @wait/next.wake@ is replaced by @yield@. 190 202 As mentioned, this benchmark may not be representative because of optimization shortcuts in @yield@. 191 The only interesting variable in this benchmark is the number of \ats per \procs, where ratios close to 1 means the ready queue(s) can be empty. 192 This scenario can put a strain on the idle-sleep handling compared to scenarios where there is plenty of work. 193 Figure~\ref{fig:yield:code} shows pseudo code for this benchmark, where the @wait/next.wake@ is replaced by @yield@. 203 The only interesting variable in this benchmark is the number of \ats per \procs, where ratios close to 1 means the ready queue(s) can be empty, which again puts a strain on the idle-sleep handling. 194 204 195 205 \begin{figure} … … 207 217 \caption[Yield Benchmark : Pseudo Code]{Yield Benchmark : Pseudo Code} 208 218 \label{fig:yield:code} 219 %\end{figure} 220 \bigskip 221 %\begin{figure} 222 \subfloat[][Throughput, 100 \ats per \proc]{ 223 \resizebox{0.5\linewidth}{!}{ 224 \input{result.yield.jax.ops.pstex_t} 225 } 226 \label{fig:yield:jax:ops} 227 } 228 \subfloat[][Throughput, 1 \ats per \proc]{ 229 \resizebox{0.5\linewidth}{!}{ 230 \input{result.yield.low.jax.ops.pstex_t} 231 } 232 \label{fig:yield:jax:low:ops} 233 } 234 235 \subfloat[][Scalability, 100 \ats per \proc]{ 236 \resizebox{0.5\linewidth}{!}{ 237 \input{result.yield.jax.ns.pstex_t} 238 } 239 \label{fig:yield:jax:ns} 240 } 241 \subfloat[][Scalability, 1 \ats per \proc]{ 242 \resizebox{0.5\linewidth}{!}{ 243 \input{result.yield.low.jax.ns.pstex_t} 244 } 245 \label{fig:yield:jax:low:ns} 246 } 247 \caption[Yield Benchmark on Intel]{Yield Benchmark on Intel\smallskip\newline Throughput and scalability as a function of \proc count, using 1 \ats per \proc. For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are extremes while the solid line is the medium.} 248 \label{fig:yield:jax} 209 249 \end{figure} 210 250 211 251 \subsection{Results} 252 253 Figures~\ref{fig:yield:jax} and~\ref{fig:yield:nasus} show the same throughput graphs as @cycle@ on Intel and AMD, respectively. 254 Note, the Y-axis on the yield graph for Intel is twice as large as the Intel cycle-graph. 255 A visual glance between the cycle and yield graphs confirms my claim that the yield benchmark is unreliable. 256 257 For the Intel architecture, Figure~\ref{fig:yield:jax}: 258 \begin{itemize} 259 \item 260 \CFA has no special handling for @yield@, but this experiment requires less synchronization than the @cycle@ experiment. 261 Hence, the @yield@ throughput and scalability graphs for both 100 and 1 cycles/tasks per processor have similar shapes to the corresponding @cycle@ graphs. 262 The only difference is sightly better performance for @yield@ because of less synchronization. 263 As for @cycle@, the cost of idle sleep also comes into play in a very significant way in Figure~\ref{fig:yield:jax:low:ns}, where the scaling is not flat. 264 \item 265 libfibre has special handling for @yield@ using the fact that the number of ready fibres does not change, and therefore, by-passing the idle-sleep mechanism entirely. 266 Additionally, when only running 1 \at per \proc, libfibre optimizes further, and forgoes the context-switch entirely. 267 Hence, libfibre behaves very differently in the cycle and yield benchmarks, with a 4 times increase in performance for 100 cycles/tasks and an 8 times increase for 1 cycle/task. 268 \item 269 Go has special handling for @yield@ by putting a yielding goroutine on a secondary global ready-queue, giving it lower priority. 270 The result is that multiple \glspl{hthrd} contend for the global queue and performance suffers drastically. 271 Hence, Go behaves very differently in the cycle and yield benchmarks, with a complete performance collapse in @yield@ for both 100 and 1 cycles/tasks. 272 \item 273 Tokio has a similar performance collapse after 16 processors, and therefore, its special @yield@ handling is probably related to a Go-like scheduler problem and/or a \CFA idle-sleep problem. 274 (I did not dig through the Rust code to ascertain the exact reason for the collapse.) 275 \end{itemize} 276 212 277 \begin{figure} 213 278 \subfloat[][Throughput, 100 \ats per \proc]{ 214 279 \resizebox{0.5\linewidth}{!}{ 215 \input{result.yield. jax.ops.pstex_t}216 } 217 \label{fig:yield: jax:ops}218 } 219 \subfloat[][Throughput, 1 \at sper \proc]{220 \resizebox{0.5\linewidth}{!}{ 221 \input{result.yield.low.jax.ops.pstex_t}222 } 223 \label{fig:yield: jax:low:ops}280 \input{result.yield.nasus.ops.pstex_t} 281 } 282 \label{fig:yield:nasus:ops} 283 } 284 \subfloat[][Throughput, 1 \at per \proc]{ 285 \resizebox{0.5\linewidth}{!}{ 286 \input{result.yield.low.nasus.ops.pstex_t} 287 } 288 \label{fig:yield:nasus:low:ops} 224 289 } 225 290 226 291 \subfloat[][Scalability, 100 \ats per \proc]{ 227 292 \resizebox{0.5\linewidth}{!}{ 228 \input{result.yield.jax.ns.pstex_t}229 }230 \label{fig:yield:jax:ns}231 }232 \subfloat[][Scalability, 1 \ats per \proc]{233 \resizebox{0.5\linewidth}{!}{234 \input{result.yield.low.jax.ns.pstex_t}235 }236 \label{fig:yield:jax:low:ns}237 }238 \caption[Yield Benchmark on Intel]{Yield Benchmark on Intel\smallskip\newline Throughput and Scalability as a function of \proc count, using 1 \ats per \proc. For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.}239 \label{fig:yield:jax}240 \end{figure}241 242 \begin{figure}243 \subfloat[][Throughput, 100 \ats per \proc]{244 \resizebox{0.5\linewidth}{!}{245 \input{result.yield.nasus.ops.pstex_t}246 }247 \label{fig:yield:nasus:ops}248 }249 \subfloat[][Throughput, 1 \at per \proc]{250 \resizebox{0.5\linewidth}{!}{251 \input{result.yield.low.nasus.ops.pstex_t}252 }253 \label{fig:yield:nasus:low:ops}254 }255 256 \subfloat[][Scalability, 100 \ats per \proc]{257 \resizebox{0.5\linewidth}{!}{258 293 \input{result.yield.nasus.ns.pstex_t} 259 294 } … … 266 301 \label{fig:yield:nasus:low:ns} 267 302 } 268 \caption[Yield Benchmark on AMD]{Yield Benchmark on AMD\smallskip\newline Throughput and Scalability as a function of \proc count, using 1 \ats per \proc. For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.}303 \caption[Yield Benchmark on AMD]{Yield Benchmark on AMD\smallskip\newline Throughput and scalability as a function of \proc count, using 1 \ats per \proc. For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are extremes while the solid line is the medium.} 269 304 \label{fig:yield:nasus} 270 305 \end{figure} 271 Figure~\ref{fig:yield:jax} shows the throughput as a function of \proc count on Intel. 272 It is fairly obvious why I claim this benchmark is more artificial. 273 The throughput is dominated by the mechanism used to handle the @yield@. 274 \CFA does not have special handling for @yield@ but the experiment requires less synchronization. 275 As a result achieves better performance than the cycle benchmark, but still comparable. 276 277 When the number of \ats is reduce to 1 per \proc, the cost of idle sleep also comes into play in a very significant way. 278 If anything causes a \at migration, where two \ats end-up on the same ready-queue, work-stealing will start occuring and could cause several \ats to shuffle around. 279 In the process, several \procs can go to sleep transiently if they fail to find where the \ats were shuffled to. 280 In \CFA, spurious bursts of latency can trick a \proc into helping, triggering this effect. 281 However, since user-level threading with equal number of \ats and \procs is a somewhat degenerate case, especially when context-switching very often, this result is not particularly meaningful and is only included for completness. 282 283 Libfibre uses the fact that @yield@ doesn't change the number of ready fibres and by-passes the idle-sleep mechanism entirely, producing significantly better throughput. 284 Additionally, when only running 1 \at per \proc, libfibre optimizes further and forgoes the context-switch entirely. 285 This results in incredible performance results comparing to the other runtimes. 286 287 In stark contrast with libfibre, Go puts yielding goroutines on a secondary global ready-queue, giving them lower priority. 288 The result is that multiple \glspl{hthrd} contend for the global queue and performance suffers drastically. 289 Based on the scalability, Tokio obtains the similarly poor performance and therefore it is likely it handles @yield@ in a similar fashion. 290 However, it must be doing something different since it does scale at low \proc count. 291 292 Again, Figure~\ref{fig:yield:nasus} show effectively the same story happening on AMD as it does on Intel. 293 \CFA fairs slightly better with many \ats per \proc, but the performance is satisfactory on both architectures. 294 295 Since \CFA obtains the same satisfactory performance as the previous benchmark this is still a success, albeit a less meaningful one. 306 307 For the AMD, Figure~\ref{fig:yield:nasus}, the results show the same story as on the Intel, with slightly increased jitter. 308 Also, some transition points on the X-axis differ because of the architectures, like at 16 versus 24 processors. 309 310 It is difficult to draw conclusions for this benchmark when runtime system treat @yield@ so differently. 311 The win for \CFA is its consistency between the cycle and yield benchmarks making it simpler for programmers to use and understand, \ie it the \CFA semantics matches with programmer intuition.. 296 312 297 313 … … 300 316 In these benchmarks, \ats can be easily partitioned over the different \procs upfront and none of the \ats communicate with each other. 301 317 302 The Churn benchmark represents more chaotic executions, where there is more communication among \ats but no apparent relation between the last \proc on which a \at ran and blocked,and the \proc that subsequently unblocks it.303 With processor-specific ready-queues, when a \at is unblocked by a different \proc that means the unblocking \proc must either ``steal'' the \at from another processor or placeit on a remote queue.304 This enqueuing results in either contention on the remote queue and/or \glspl{rmr} on the \at data structure.318 The Churn benchmark represents more chaotic executions, where there is more communication among \ats but no relationship between the last \proc on which a \at ran and blocked and the \proc that subsequently unblocks it. 319 With processor-specific ready-queues, when a \at is unblocked by a different \proc that means the unblocking \proc must either ``steal'' the \at from another processor or find it on a remote queue. 320 This dequeuing results in either contention on the remote queue and/or \glspl{rmr} on the \at data structure. 305 321 In either case, this benchmark aims to measure how well each scheduler handles these cases, since both cases can lead to performance degradation if not handled correctly. 306 322 … … 328 344 \caption[Churn Benchmark : Pseudo Code]{Churn Benchmark : Pseudo Code} 329 345 \label{fig:churn:code} 346 %\end{figure} 347 \bigskip 348 %\begin{figure} 349 \subfloat[][Throughput, 100 \ats per \proc]{ 350 \resizebox{0.5\linewidth}{!}{ 351 \input{result.churn.jax.ops.pstex_t} 352 } 353 \label{fig:churn:jax:ops} 354 } 355 \subfloat[][Throughput, 2 \ats per \proc]{ 356 \resizebox{0.5\linewidth}{!}{ 357 \input{result.churn.low.jax.ops.pstex_t} 358 } 359 \label{fig:churn:jax:low:ops} 360 } 361 362 \subfloat[][Latency, 100 \ats per \proc]{ 363 \resizebox{0.5\linewidth}{!}{ 364 \input{result.churn.jax.ns.pstex_t} 365 } 366 \label{fig:churn:jax:ns} 367 } 368 \subfloat[][Latency, 2 \ats per \proc]{ 369 \resizebox{0.5\linewidth}{!}{ 370 \input{result.churn.low.jax.ns.pstex_t} 371 } 372 \label{fig:churn:jax:low:ns} 373 } 374 \caption[Churn Benchmark on Intel]{\centering Churn Benchmark on Intel\smallskip\newline Throughput and latency of the Churn on the benchmark on the Intel machine. For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are extremes while the solid line is the medium.} 375 \label{fig:churn:jax} 330 376 \end{figure} 331 377 332 378 \subsection{Results} 333 \begin{figure} 334 \subfloat[][Throughput, 100 \ats per \proc]{ 335 \resizebox{0.5\linewidth}{!}{ 336 \input{result.churn.jax.ops.pstex_t} 337 } 338 \label{fig:churn:jax:ops} 339 } 340 \subfloat[][Throughput, 2 \ats per \proc]{ 341 \resizebox{0.5\linewidth}{!}{ 342 \input{result.churn.low.jax.ops.pstex_t} 343 } 344 \label{fig:churn:jax:low:ops} 345 } 346 347 \subfloat[][Latency, 100 \ats per \proc]{ 348 \resizebox{0.5\linewidth}{!}{ 349 \input{result.churn.jax.ns.pstex_t} 350 } 351 \label{fig:churn:jax:ns} 352 } 353 \subfloat[][Latency, 2 \ats per \proc]{ 354 \resizebox{0.5\linewidth}{!}{ 355 \input{result.churn.low.jax.ns.pstex_t} 356 } 357 \label{fig:churn:jax:low:ns} 358 } 359 \caption[Churn Benchmark on Intel]{\centering Churn Benchmark on Intel\smallskip\newline Throughput and latency of the Churn on the benchmark on the Intel machine. For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.} 360 \label{fig:churn:jax} 361 \end{figure} 362 363 \begin{figure} 364 \subfloat[][Throughput, 100 \ats per \proc]{ 365 \resizebox{0.5\linewidth}{!}{ 366 \input{result.churn.nasus.ops.pstex_t} 367 } 368 \label{fig:churn:nasus:ops} 369 } 370 \subfloat[][Throughput, 2 \ats per \proc]{ 371 \resizebox{0.5\linewidth}{!}{ 372 \input{result.churn.low.nasus.ops.pstex_t} 373 } 374 \label{fig:churn:nasus:low:ops} 375 } 376 377 \subfloat[][Latency, 100 \ats per \proc]{ 378 \resizebox{0.5\linewidth}{!}{ 379 \input{result.churn.nasus.ns.pstex_t} 380 } 381 \label{fig:churn:nasus:ns} 382 } 383 \subfloat[][Latency, 2 \ats per \proc]{ 384 \resizebox{0.5\linewidth}{!}{ 385 \input{result.churn.low.nasus.ns.pstex_t} 386 } 387 \label{fig:churn:nasus:low:ns} 388 } 389 \caption[Churn Benchmark on AMD]{\centering Churn Benchmark on AMD\smallskip\newline Throughput and latency of the Churn on the benchmark on the AMD machine. 390 For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.} 391 \label{fig:churn:nasus} 392 \end{figure} 393 Figure~\ref{fig:churn:jax} and Figure~\ref{fig:churn:nasus} show the throughput as a function of \proc count on Intel and AMD respectively. 394 It uses the same representation as the previous benchmark : 15 runs where the dashed line show the extremums and the solid line the median. 379 Figures~\ref{fig:churn:jax} and Figure~\ref{fig:churn:nasus} show the throughput as a function of \proc count on Intel and AMD respectively. 380 It uses the same representation as the previous benchmark : 15 runs where the dashed line show the extremes and the solid line the median. 395 381 The performance cost of crossing the cache boundaries is still visible at the same \proc count. 396 However, this benchmark has performance dominated by the cache traffic as \proc are constantly accessing the each other's data.382 However, this benchmark has performance dominated by the cache traffic as \proc are constantly accessing the each other's data. 397 383 Scalability is notably worst than the previous benchmarks since there is inherently more communication between processors. 398 384 Indeed, once the number of \glspl{hthrd} goes beyond a single socket, performance ceases to improve. 399 385 An interesting aspect to note here is that the runtimes differ in how they handle this situation. 400 386 Indeed, when a \proc unparks a \at that was last run on a different \proc, the \at could be appended to the ready-queue local \proc or to the ready-queue of the remote \proc, which previously ran the \at. 401 \CFA, tokio and Go all use the approach of unparking to the local \proc while Libfibre unparks to the remote \proc.387 \CFA, Tokio and Go all use the approach of unparking to the local \proc while Libfibre unparks to the remote \proc. 402 388 In this particular benchmark, the inherent chaos of the benchmark in addition to small memory footprint means neither approach wins over the other. 389 390 \begin{figure} 391 \subfloat[][Throughput, 100 \ats per \proc]{ 392 \resizebox{0.5\linewidth}{!}{ 393 \input{result.churn.nasus.ops.pstex_t} 394 } 395 \label{fig:churn:nasus:ops} 396 } 397 \subfloat[][Throughput, 2 \ats per \proc]{ 398 \resizebox{0.5\linewidth}{!}{ 399 \input{result.churn.low.nasus.ops.pstex_t} 400 } 401 \label{fig:churn:nasus:low:ops} 402 } 403 404 \subfloat[][Latency, 100 \ats per \proc]{ 405 \resizebox{0.5\linewidth}{!}{ 406 \input{result.churn.nasus.ns.pstex_t} 407 } 408 \label{fig:churn:nasus:ns} 409 } 410 \subfloat[][Latency, 2 \ats per \proc]{ 411 \resizebox{0.5\linewidth}{!}{ 412 \input{result.churn.low.nasus.ns.pstex_t} 413 } 414 \label{fig:churn:nasus:low:ns} 415 } 416 \caption[Churn Benchmark on AMD]{\centering Churn Benchmark on AMD\smallskip\newline Throughput and latency of the Churn on the benchmark on the AMD machine. 417 For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are extremes while the solid line is the medium.} 418 \label{fig:churn:nasus} 419 \end{figure} 403 420 404 421 Like for the cycle benchmark, here all runtimes achieve fairly similar performance. … … 406 423 Beyond that performance starts to suffer from increased caching costs. 407 424 408 Indeed on Figures~\ref{fig:churn:jax:ops} and \ref{fig:churn:jax:ns} show that with 1 and 100 \ats per \proc, \CFA, libfibre, Go and tokio achieve effectively equivalent performance for most \proc count.425 Indeed on Figures~\ref{fig:churn:jax:ops} and \ref{fig:churn:jax:ns} show that with 1 and 100 \ats per \proc, \CFA, libfibre, Go and Tokio achieve effectively equivalent performance for most \proc count. 409 426 410 427 However, Figure~\ref{fig:churn:nasus} again shows a somewhat different story on AMD. 411 While \CFA, libfibre, and tokio achieve effectively equivalent performance for most \proc count, Go starts with better scaling at very low \proc counts but then performance quickly plateaus, resulting in worse performance at higher \proc counts.428 While \CFA, libfibre, and Tokio achieve effectively equivalent performance for most \proc count, Go starts with better scaling at very low \proc counts but then performance quickly plateaus, resulting in worse performance at higher \proc counts. 412 429 This performance difference is visible at both high and low \at counts. 413 430 … … 420 437 As second possible explanation is that Go may sometimes use the heap when allocating variables based on the result of escape analysis of the code. 421 438 It is possible that variables that should be placed on the stack are placed on the heap. 422 This could cause extra pointer chasing in the benchmark, height ning locality effects.439 This could cause extra pointer chasing in the benchmark, heightening locality effects. 423 440 Depending on how the heap is structure, this could also lead to false sharing. 424 441 425 442 The objective of this benchmark is to demonstrate that unparking \ats from remote \procs do not cause too much contention on the local queues. 426 Indeed, the fact all runtimes achieve some scaling at lower \proc count demon trate that migrations do not need to be serialized.443 Indeed, the fact all runtimes achieve some scaling at lower \proc count demonstrate that migrations do not need to be serialized. 427 444 Again these result demonstrate \CFA achieves satisfactory performance. 428 445 429 446 \section{Locality} 430 \begin{figure} 431 \begin{cfa} 447 448 \begin{figure} 449 \newsavebox{\myboxA} 450 \newsavebox{\myboxB} 451 452 \begin{lrbox}{\myboxA} 453 \begin{cfa}[tabsize=3] 432 454 Thread.main() { 433 455 count := 0 … … 436 458 // go through the array 437 459 @work( a )@ 460 438 461 spots[r].V() 439 462 spots[r].P() … … 444 467 } 445 468 \end{cfa} 446 \begin{cfa} 469 \end{lrbox} 470 471 \begin{lrbox}{\myboxB} 472 \begin{cfa}[tabsize=3] 447 473 Thread.main() { 448 474 count := 0 … … 460 486 } 461 487 \end{cfa} 488 \end{lrbox} 489 490 \subfloat[Thread$_1$]{\label{f:CFibonacci}\usebox\myboxA} 491 \hspace{3pt} 492 \vrule 493 \hspace{3pt} 494 \subfloat[Thread$_2$]{\label{f:CFAFibonacciGen}\usebox\myboxB} 495 462 496 \caption[Locality Benchmark : Pseudo Code]{Locality Benchmark : Pseudo Code} 463 497 \label{fig:locality:code} 464 498 \end{figure} 465 As mentionned in the churn benchmark, when unparking a \at, it is possible to either unpark to the local or remote ready-queue. 499 500 As mentioned in the churn benchmark, when unparking a \at, it is possible to either unpark to the local or remote ready-queue. 466 501 \footnote{It is also possible to unpark to a third unrelated ready-queue, but without additional knowledge about the situation, there is little to suggest this would not degrade performance.} 467 502 The locality experiment includes two variations of the churn benchmark, where an array of data is added. 468 503 In both variations, before @V@ing the semaphore, each \at increment random cells inside the array. 469 504 The @share@ variation then passes the array to the shadow-queue of the semaphore, transferring ownership of the array to the woken thread. 470 In the @noshare@ variation the array is not passed on and each thread contin ously accesses its private array.505 In the @noshare@ variation the array is not passed on and each thread continuously accesses its private array. 471 506 472 507 The objective here is to highlight the different decision made by the runtime when unparking. … … 480 515 481 516 \subsection{Results} 517 482 518 \begin{figure} 483 519 \subfloat[][Throughput share]{ … … 506 542 \label{fig:locality:jax:noshare:ns} 507 543 } 508 \caption[Locality Benchmark on Intel]{Locality Benchmark on Intel\smallskip\newline Throughput and Scalability as a function of \proc count. For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.}544 \caption[Locality Benchmark on Intel]{Locality Benchmark on Intel\smallskip\newline Throughput and scalability as a function of \proc count. For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are extremes while the solid line is the medium.} 509 545 \label{fig:locality:jax} 510 546 \end{figure} … … 535 571 \label{fig:locality:nasus:noshare:ns} 536 572 } 537 \caption[Locality Benchmark on AMD]{Locality Benchmark on AMD\smallskip\newline Throughput and Scalability as a function of \proc count. For Throughput higher is better, for Scalability lower is better. Each series represent 15 independent runs, the dotted lines are extremums while the solid line is the medium.}573 \caption[Locality Benchmark on AMD]{Locality Benchmark on AMD\smallskip\newline Throughput and scalability as a function of \proc count. For throughput, higher is better, for scalability, lower is better. Each series represent 15 independent runs, the dotted lines are extremes while the solid line is the medium.} 538 574 \label{fig:locality:nasus} 539 575 \end{figure} 540 576 541 Figure ~\ref{fig:locality:jax} and \ref{fig:locality:nasus} shows the results on Intel and AMD respectively.577 Figures~\ref{fig:locality:jax} and \ref{fig:locality:nasus} shows the results on Intel and AMD respectively. 542 578 In both cases, the graphs on the left column show the results for the @share@ variation and the graphs on the right column show the results for the @noshare@. 543 579 544 580 On Intel, Figure~\ref{fig:locality:jax} shows Go trailing behind the 3 other runtimes. 545 On the left of the figure showing the results for the shared variation, where \CFA and tokio slightly outperform libfibre as expected.546 And correspondingly on the right, we see the expected performance inversion where libfibre now outperforms \CFA and tokio.547 Otherwise the results are similar to the churn benchmark, with lower through tput due to the array processing.581 On the left of the figure showing the results for the shared variation, where \CFA and Tokio slightly outperform libfibre as expected. 582 And correspondingly on the right, we see the expected performance inversion where libfibre now outperforms \CFA and Tokio. 583 Otherwise the results are similar to the churn benchmark, with lower throughput due to the array processing. 548 584 Presumably the reason why Go trails behind are the same as in Figure~\ref{fig:churn:nasus}. 549 585 550 586 Figure~\ref{fig:locality:nasus} shows the same experiment on AMD. 551 587 \todo{why is cfa slower?} 552 Again, we see the same story, where tokio and libfibre swap places and Go trails behind.588 Again, we see the same story, where Tokio and libfibre swap places and Go trails behind. 553 589 554 590 \section{Transfer} … … 572 608 In both flavours, the experiment effectively measures how long it takes for all \ats to run once after a given synchronization point. 573 609 In an ideal scenario where the scheduler is strictly FIFO, every thread would run once after the synchronization and therefore the delay between leaders would be given by: 574 $ \frac{CSL + SL}{NP - 1}$, where $CSL$ is the context switch latency, $SL$ is the cost for enqueu ing and dequeuing a \at and $NP$ is the number of \procs.610 $ \frac{CSL + SL}{NP - 1}$, where $CSL$ is the context switch latency, $SL$ is the cost for enqueueing and dequeuing a \at and $NP$ is the number of \procs. 575 611 However, if the scheduler allows \ats to run many times before other \ats are able to run once, this delay will increase. 576 612 The semaphore version is an approximation of the strictly FIFO scheduling, where none of the \ats \emph{attempt} to run more than once. 577 613 The benchmark effectively provides the fairness guarantee in this case. 578 In the yielding version however, the benchmark provides no such guarantee, which means the scheduler has full respons ability and any unfairness will be measurable.614 In the yielding version however, the benchmark provides no such guarantee, which means the scheduler has full responsibility and any unfairness will be measurable. 579 615 580 616 While this is a fairly artificial scenario, it requires only a few simple pieces. … … 634 670 libfibre & 127 $\mu$s & ~33.5 ms & DNC & DNC & 156 $\mu$s & ~36.7 ms & DNC & DNC \\ 635 671 Go & 106 $\mu$s & ~64.0 ms & 24.6 ms & 74.3 ms & 271 $\mu$s & 121.6 ms & ~~1.21~ms & 117.4 ms \\ 636 tokio & 289 $\mu$s & 180.6 ms & DNC & DNC & 157 $\mu$s & 111.0 ms & DNC & DNC672 Tokio & 289 $\mu$s & 180.6 ms & DNC & DNC & 157 $\mu$s & 111.0 ms & DNC & DNC 637 673 \end{tabular} 638 674 \end{centering} 639 \caption[Transfer Benchmark on Intel and AMD]{Transfer Benchmark on Intel and AMD\smallskip\newline Average measurement of how long it takes for all \ats to acknowledge the leader \at. DNC stands for ``did not complete'', meaning that after 5 seconds of a new leader being decided, some \ats still had not acknowledged the new leader. 675 \caption[Transfer Benchmark on Intel and AMD]{Transfer Benchmark on Intel and AMD\smallskip\newline Average measurement of how long it takes for all \ats to acknowledge the leader \at. DNC stands for ``did not complete'', meaning that after 5 seconds of a new leader being decided, some \ats still had not acknowledged the new leader.} 640 676 \label{fig:transfer:res} 641 677 \end{figure} 642 Figure~\ref{fig:transfer:res} shows the result for the transfer benchmark with 2 \procs and all \procs, where each experiement runs 100 \at per \proc. 678 679 Figure~\ref{fig:transfer:res} shows the result for the transfer benchmark with 2 \procs and all \procs, where each experiment runs 100 \at per \proc. 643 680 Note that the results here are only meaningful as a coarse measurement of fairness, beyond which small cost differences in the runtime and concurrent primitives begin to matter. 644 As such, data points that are the on the same order of magnitude as each other should be basically considered equal.645 The takeaway of this experi ement is the presence of very large differences.681 As such, data points that are the on the same order of magnitude as each other should be basically considered equal. 682 The takeaway of this experiment is the presence of very large differences. 646 683 The semaphore variation is denoted ``Park'', where the number of \ats dwindles down as the new leader is acknowledged. 647 684 The yielding variation is denoted ``Yield''. 648 The experi ement was only run for the extremums of the number of cores since the scaling per core behaves like previous experiements.685 The experiment was only run for the extremes of the number of cores since the scaling per core behaves like previous experiments. 649 686 This experiments clearly demonstrate that while the other runtimes achieve similar performance in previous benchmarks, here \CFA achieves significantly better fairness. 650 687 The semaphore variation serves as a control group, where all runtimes are expected to transfer leadership fairly quickly. … … 652 689 Figure~\ref{fig:transfer:res} shows that while Go and Tokio are slower, all runtime achieve decent latency. 653 690 However, the yielding variation shows an entirely different picture. 654 Since libfibre and tokio have a traditional work-stealing scheduler, \procs that have \ats on their local queues will never steal from other \procs.655 The result is that the experi ement simply does not complete for these runtime.691 Since libfibre and Tokio have a traditional work-stealing scheduler, \procs that have \ats on their local queues will never steal from other \procs. 692 The result is that the experiment simply does not complete for these runtime. 656 693 Without \procs stealing from the \proc running the leader, the experiment will simply never terminate. 657 Go manages to complete the experi ement because it adds preemption on top of classic work-stealing.694 Go manages to complete the experiment because it adds preemption on top of classic work-stealing. 658 695 However, since preemption is fairly costly it achieves significantly worst performance. 659 696 In contrast, \CFA achieves equivalent performance in both variations, demonstrating very good fairness. -
doc/theses/thierry_delisle_PhD/thesis/text/existing.tex
r683cc13 r36cc24a 50 50 It can therefore be desirable for schedulers to support \ats with identical priorities and/or automatically setting and adjusting priorities for \ats. 51 51 Most common operating systems use some variant on priorities with overlaps and dynamic priority adjustments. 52 For example, Microsoft Windows uses a pair of priorities 53 \cite{win:priority}, one specified by users out of ten possible options and one adjusted by the system. 52 For example, Microsoft Windows uses a pair of priorities~\cite{win:priority}, one specified by users out of ten possible options and one adjusted by the system. 54 53 55 54 \subsection{Uninformed and Self-Informed Dynamic Schedulers} … … 137 136 The scheduler may also temporarily adjust priorities after certain effects like the completion of I/O requests. 138 137 139 In~\cite{russinovich2009windows}, Chapter 1 section ``Processes, Threads, and Jobs'' discusses the scheduling policy more in depth.140 Multicore scheduling is based on a combination of priorities ,preferred \proc.141 Each \at is assigned an \newterm{ideal} \proc using a round-robin policy.142 \Gls {at} are distributed among the \procs according to their priority, preferring to match \ats to their ideal \proc and then to the last \proc they ran on.143 This is similar to a variation of work stealing, where the stealing \proc restore the \at to its original \proc after running it, but with priorities added onto the mix.138 In~\cite{russinovich2009windows}, Chapter 1 section ``Processes, Threads, and Jobs''\todo{Look up section number.} discusses the scheduling policy more in depth. 139 Multicore scheduling is based on a combination of priorities and preferred \proc. 140 Each \at is assigned an initial processor using a round-robin policy, called the \at's \newterm{ideal} \proc. 141 \Glspl{at} are distributed among the \procs according to their priority, preferring to match \ats to their ideal \proc and then to the last \proc they ran on. 142 This approach is a variation of work stealing, where the stealing \proc restore the \at to its original \proc after running it, but mixed with priorities. 144 143 145 144 \paragraph{Apple OS X} … … 203 202 204 203 \paragraph{Grand Central Dispatch} 205 An Apple \cite{apple:gcd} API that offers task parallelism~\cite{wiki:taskparallel}.204 An Apple~\cite{apple:gcd} API that offers task parallelism~\cite{wiki:taskparallel}. 206 205 Its distinctive aspect is multiple ``Dispatch Queues'', some of which are created by programmers. 207 206 Each queue has its own local ordering guarantees, \eg \ats on queue $A$ are executed in \emph{FIFO} order. 208 207 209 While the documentation only gives limited insight into the scheduling and load balancing approach, \cite{apple:gcd2} suggests an approach fairly classic; 210 Where each \proc has a queue of \newterm{blocks} to run, effectively \ats, and they drain their respective queues in \glsxtrshort{fifo}. 211 They seem to add the concept of dependent queues with clear ordering, where a executing a block ends-up scheduling more blocks. 212 In terms of semantics, these Dispatch Queues seem to be very similar to Intel\textregistered ~TBB @execute()@ and predecessor semantics. 208 While the documentation only gives limited insight into the scheduling and load balancing approach, \cite{apple:gcd2} suggests a fairly classic approach. 209 Each \proc has a queue of \ats to run, called \newterm{blocks}, which are drained in \glsxtrshort{fifo}. 210 \todo{update: They seem to add the concept of dependent queues with clear ordering, where executing a block ends-up scheduling more blocks. 211 In terms of semantics, these Dispatch Queues seem to be very similar to Intel\textregistered ~TBB \lstinline{execute()} and predecessor semantics.} 212 213 213 214 214 \paragraph{LibFibre} -
doc/theses/thierry_delisle_PhD/thesis/text/io.tex
r683cc13 r36cc24a 141 141 In the worst case, where all \glspl{thrd} are consistently blocking on \io, it devolves into 1-to-1 threading. 142 142 However, regardless of the frequency of \io operations, it achieves the fundamental goal of not blocking \glspl{proc} when \glspl{thrd} are ready to run. 143 This approach is used by languages like Go \cite{GITHUB:go}, frameworks like libuv\cite{libuv}, and web servers like Apache~\cite{apache} and Nginx~\cite{nginx}, since it has the advantage that it can easily be used across multiple operating systems.143 This approach is used by languages like Go~\cite{GITHUB:go}, frameworks like libuv~\cite{libuv}, and web servers like Apache~\cite{apache} and NGINX~\cite{nginx}, since it has the advantage that it can easily be used across multiple operating systems. 144 144 This advantage is especially relevant for languages like Go, which offer a homogeneous \glsxtrshort{api} across all platforms. 145 145 As opposed to C, which has a very limited standard api for \io, \eg, the C standard library has no networking. … … 151 151 152 152 For this project, I selected @io_uring@, in large parts because of its generality. 153 While @epoll@ has been shown to be a good solution for socket \io (\cite{ DBLP:journals/pomacs/KarstenB20}), @io_uring@'s transparent support for files, pipes, and more complex operations, like @splice@ and @tee@, make it a better choice as the foundation for a general \io subsystem.153 While @epoll@ has been shown to be a good solution for socket \io (\cite{Karsten20}), @io_uring@'s transparent support for files, pipes, and more complex operations, like @splice@ and @tee@, make it a better choice as the foundation for a general \io subsystem. 154 154 155 155 \section{Event-Engine}
Note: See TracChangeset
for help on using the changeset viewer.