`scan-build` is part of our regular CI suite. Other static analyzers are not.
It can be useful to look at additional static analyzers once in a while (and we do), but it's not a good idea to multiply the nb of analyzers run continuously at each commit and PR. The reasons are :
- Static analyzers are full of false positive. The signal to noise ratio is actually pretty low.
- A good CI policy is "zero-warning tolerance". That means that all issues must be solved, including false positives. This quickly becomes a tedious workload.
- Multiple static analyzers will feature multiple kind of false positives, sometimes applying to the same code but in different ways leading to :
+ torteous code, trying to please multiple constraints, hurting readability and therefore maintenance. Sometimes, such complexity introduce other more subtle bugs, that are just out of scope of the analyzers.
+ sometimes, these constraints are mutually exclusive : if one try to solve one, the other static analyzer will complain, they can't be both happy at the same time.
- As if that was not enough, the list of false positives change with each version. It's hard enough to follow one static analyzer, but multiple ones with their own update agenda, this quickly becomes a massive velocity reducer.
This is different from running a static analyzer once in a while, looking at the output, and __cherry picking__ a few warnings that seem helpful, either because they detected a genuine risk of bug, or because it helps expressing the code in a way which is more readable or more difficult to misuse. These kind of reports can be useful, and are accepted.
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.
