From e7d2391e9a75154657183b049b69a7a2effa9724 Mon Sep 17 00:00:00 2001 From: Bimba Shrestha Date: Mon, 11 May 2020 20:55:01 -0500 Subject: [PATCH] [doc] measuring performance docs (#2117) * performance measuring docs * spelling * combining advanced and simple section * zstd benchmark title change --- CONTRIBUTING.md | 204 +++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 202 insertions(+), 2 deletions(-) diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 23d62c09..637e3718 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -44,7 +44,7 @@ Our contribution process works in three main stages: git pull https://github.com/facebook/zstd dev git push origin dev ``` - * Topic and deveopment: + * Topic and development: * Make a new branch on your fork about the topic you're developing for ``` # branch names should be consise but sufficiently informative @@ -80,7 +80,7 @@ Our contribution process works in three main stages: as the destination. * Examine the diff presented between the two branches to make sure there is nothing unexpected. * Write a good pull request description: - * While there is no strict template that our contributers follow, we would like them to + * While there is no strict template that our contributors follow, we would like them to sufficiently summarize and motivate the changes they are proposing. We recommend all pull requests, at least indirectly, address the following points. * Is this pull request important and why? @@ -126,6 +126,206 @@ just `contrib/largeNbDicts` and nothing else, you can run: scan-build make -C contrib/largeNbDicts largeNbDicts ``` +## Performance +Performance is extremely important for zstd and we only merge pull requests whose performance +landscape and corresponding trade-offs have been adequately analyzed, reproduced, and presented. +This high bar for performance means that every PR which has the potential to +impact performance takes a very long time for us to properly review. That being said, we +always welcome contributions to improve performance (or worsen performance for the trade-off of +something else). Please keep the following in mind before submitting a performance related PR: + +1. Zstd isn't as old as gzip but it has been around for time now and its evolution is +very well documented via past Github issues and pull requests. It may be the case that your +particular performance optimization has already been considered in the past. Please take some +time to search through old issues and pull requests using keywords specific to your +would-be PR. Of course, just because a topic has already been discussed (and perhaps rejected +on some grounds) in the past, doesn't mean it isn't worth bringing up again. But even in that case, +it will be helpful for you to have context from that topic's history before contributing. +2. The distinction between noise and actual performance gains can unfortunately be very subtle +especially when microbenchmarking extremely small wins or losses. The only remedy to getting +something subtle merged is extensive benchmarking. You will be doing us a great favor if you +take the time to run extensive, long-duration, and potentially cross-(os, platform, process, etc) +benchmarks on your end before submitting a PR. Of course, you will not be able to benchmark +your changes on every single processor and os out there (and neither will we) but do that best +you can:) We've adding some things to think about when benchmarking below in the Benchmarking +Performance section which might be helpful for you. +3. Optimizing performance for a certain OS, processor vendor, compiler, or network system is a perfectly +legitimate thing to do as long as it does not harm the overall performance health of Zstd. +This is a hard balance to strike but please keep in mind other aspects of Zstd when +submitting changes that are clang-specific, windows-specific, etc. + +## Benchmarking Performance +Performance microbenchmarking is a tricky subject but also essential for Zstd. We value empirical +testing over theoretical speculation. This guide it not perfect but for most scenarios, it +is a good place to start. + +### Stability +Unfortunately, the most important aspect in being able to benchmark reliably is to have a stable +benchmarking machine. A virtual machine, a machine with shared resources, or your laptop +will typically not be stable enough to obtain reliable benchmark results. If you can get your +hands on a desktop, this is usually a better scenario. + +Of course, benchmarking can be done on non-hyper-stable machines as well. You will just have to +do a little more work to ensure that you are in fact measuring the changes you've made not and +noise. Here are some things you can do to make your benchmarks more stable: + +1. The most simple thing you can do to drastically improve the stability of your benchmark is +to run it multiple times and then aggregate the results of those runs. As a general rule of +thumb, the smaller the change you are trying to measure, the more samples of benchmark runs +you will have to aggregate over to get reliable results. Here are some additional things to keep in +mind when running multiple trials: + * How you aggregate your samples are important. You might be tempted to use the mean of your + results. While this is certainly going to be a more stable number than a raw single sample + benchmark number, you might have more luck by taking the median. The mean is not robust to + outliers whereas the median is. Better still, you could simply take the fastest speed your + benchmark achieved on each run since that is likely the fastest your process will be + capable of running your code. In our experience, this (aggregating by just taking the sample + with the fastest running time) has been the most stable approach. + * The more samples you have, the more stable your benchmarks should be. You can verify + your improved stability by looking at the size of your confidence intervals as you + increase your sample count. These should get smaller and smaller. Eventually hopefully + smaller than the performance win you are expecting. + * Most processors will take some time to get `hot` when running anything. The observations + you collect during that time period will very different from the true performance number. Having + a very large number of sample will help alleviate this problem slightly but you can also + address is directly by simply not including the first `n` iterations of your benchmark in + your aggregations. You can determine `n` by simply looking at the results from each iteration + and then hand picking a good threshold after which the variance in results seems to stabilize. +2. You cannot really get reliable benchmarks if your host machine is simultaneously running +another cpu/memory-intensive application in the background. If you are running benchmarks on your +personal laptop for instance, you should close all applications (including your code editor and +browser) before running your benchmarks. You might also have invisible background applications +running. You can see what these are by looking at either Activity Monitor on Mac or Task Manager +on Windows. You will get more stable benchmark results of you end those processes as well. + * If you have multiple cores, you can even run your benchmark on a reserved core to prevent + pollution from other OS and user processes. There are a number of ways to do this depending + on your OS: + * On linux boxes, you have use https://github.com/lpechacek/cpuset. + * On Windows, you can "Set Processor Affinity" using https://www.thewindowsclub.com/processor-affinity-windows + * On Mac, you can try to use their dedicated affinity API https://developer.apple.com/library/archive/releasenotes/Performance/RN-AffinityAPI/#//apple_ref/doc/uid/TP40006635-CH1-DontLinkElementID_2 +3. To benchmark, you will likely end up writing a separate c/c++ program that will link libzstd. +Dynamically linking your library will introduce some added variation (not a large amount but +definitely some). Statically linking libzstd will be more stable. Static libraries should +be enabled by default when building zstd. +4. Use a profiler with a good high resolution timer. See the section below on profiling for +details on this. +5. Disable frequency scaling, turbo boost and address space randomization (this will vary by OS) +6. Try to avoid storage. On some systems you can use tmpfs. Putting the program, inputs and outputs on +tmpfs avoids touching a real storage system, which can have a pretty big variability. + +Also check our LLVM's guide on benchmarking here: https://llvm.org/docs/Benchmarking.html + +### Zstd benchmark +The fastest signal you can get regarding your performance changes is via the in-build zstd cli +bench option. You can run Zstd as you typically would for your scenario using some set of options +and then additionally also specify the `-b#` option. Doing this will run our benchmarking pipeline +for that options you have just provided. If you want to look at the internals of how this +benchmarking script works, you can check out programs/benchzstd.c + +For example: say you have made a change that you believe improves the speed of zstd level 1. The +very first thing you should use to asses whether you actually achieved any sort of improvement +is `zstd -b`. You might try to do something like this. Note: you can use the `-i` option to +specify a running time for your benchmark in seconds (default is 3 seconds). +Usually, the longer the running time, the more stable your results will be. + +``` +$ git checkout +$ make && cp zstd zstd-old +$ git checkout +$ make && cp zstd zstd-new +$ zstd-old -i5 -b1 + 1 : 8990 -> 3992 (2.252), 302.6 MB/s , 626.4 MB/s +$ zstd-new -i5 -b1 + 1 : 8990 -> 3992 (2.252), 302.8 MB/s , 628.4 MB/s +``` + +Unless your performance win is large enough to be visible despite the intrinsic noise +on your computer, benchzstd alone will likely not be enough to validate the impact of your +changes. For example, the results of the example above indicate that effectively nothing +changed but there could be a small <3% improvement that the noise on the host machine +obscured. So unless you see a large performance win (10-15% consistently) using just +this method of evaluation will not be sufficient. + +### Profiling +There are a number of great profilers out there. We're going to briefly mention how you can +profile your code using `instruments` on mac, `perf` on linux and `visual studio profiler` +on windows. + +Say you have an idea for a change that you think will provide some good performance gains +for level 1 compression on Zstd. Typically this means, you have identified a section of +code that you think can be made to run faster. + +The first thing you will want to do is make sure that the piece of code is actually taking up +a notable amount of time to run. It is usually not worth optimzing something which accounts for less than +0.0001% of the total running time. Luckily, there are tools to help with this. +Profilers will let you see how much time your code spends inside a particular function. +If your target code snippit is only part of a function, it might be worth trying to +isolate that snippit by moving it to its own function (this is usually not necessary but +might be). + +Most profilers (including the profilers dicusssed below) will generate a call graph of +functions for you. Your goal will be to find your function of interest in this call grapch +and then inspect the time spent inside of it. You might also want to to look at the +annotated assembly which most profilers will provide you with. + +#### Instruments +We will once again consider the scenario where you think you've identified a piece of code +whose performance can be improved upon. Follow these steps to profile your code using +Instruments. + +1. Open Instruments +2. Select `Time Profiler` from the list of standard templates +3. Close all other applications except for your instruments window and your terminal +4. Run your benchmarking script from your terminal window + * You will want a benchmark that runs for at least a few seconds (5 seconds will + usually be long enough). This way the profiler will have something to work with + and you will have ample time to attach your profiler to this process:) + * I will just use benchzstd as my bencharmking script for this example: +``` +$ zstd -b1 -i5 # this will run for 5 seconds +``` +5. Once you run your benchmarking script, switch back over to instruments and attach your +process to the time profiler. You can do this by: + * Clicking on the `All Processes` drop down in the top left of the toolbar. + * Selecting your process from the dropdown. In my case, it is just going to be labled + `zstd` + * Hitting the bright red record circle button on the top left of the toolbar +6. You profiler will now start collecting metrics from your bencharking script. Once +you think you have collected enough samples (usually this is the case after 3 seconds of +recording), stop your profiler. +7. Make sure that in toolbar of the bottom window, `profile` is selected. +8. You should be able to see your call graph. + * If you don't see the call graph or an incomplete call graph, make sure you have compiled + zstd and your benchmarking scripg using debug flags. On mac and linux, this just means + you will have to supply the `-g` flag alone with your build script. You might also + have to provide the `-fno-omit-frame-pointer` flag +9. Dig down the graph to find your function call and then inspect it by double clicking +the list item. You will be able to see the annotated source code and the assembly side by +side. + +#### Perf + +This wiki has a pretty detailed tutorial on getting started working with perf so we'll +leave you to check that out of you're getting started: + +https://perf.wiki.kernel.org/index.php/Tutorial + +Some general notes on perf: +* Use `perf stat -r # ` to quickly get some relevant timing and +counter statistics. Perf uses a high resolution timer and this is likely one +of the first things your team will run when assessing your PR. +* Perf has a long list of hardware counters that can be viewed with `perf --list`. +When measuring optimizations, something worth trying is to make sure the handware +counters you expect to be impacted by your change are in fact being so. For example, +if you expect the L1 cache misses to decrease with your change, you can look at the +counter `L1-dcache-load-misses` +* Perf hardware counters will not work on a virtual machine. + +#### Visual Studio + +TODO + + ## Setting up continuous integration (CI) on your fork Zstd uses a number of different continuous integration (CI) tools to ensure that new changes are well tested before they make it to an official release. Specifically, we use the platforms