add a page on optimizations and profiling #45
Conversation
src/development/perf-benchmarking.md
Outdated
| e.g. to run it under `perf stat` or cachegrind. | ||
|
|
||
| Build and link the [stage1](https://rustc-dev-guide.rust-lang.org/building/how-to-build-and-run.html#creating-a-rustup-toolchain) | ||
| compiler as rustup toolchain and then use that to build the standalone benchmark with a modified standard library. |
There was a problem hiding this comment.
Thanks for linking this! I want to reland that PR but haven't had time.
| * [disable ASLR](https://man7.org/linux/man-pages/man8/setarch.8.html) | ||
| * [pinning](https://man7.org/linux/man-pages/man1/taskset.1.html) the benchmark process to a specific core | ||
| * [disable clock boosts](https://wiki.archlinux.org/title/CPU_frequency_scaling#Configuring_frequency_boosting), | ||
| especially on thermal-limited systems such as laptops |
There was a problem hiding this comment.
We can also add a link to something like https://github.com/JuliaCI/BenchmarkTools.jl/blob/master/docs/src/linuxtips.md.
There was a problem hiding this comment.
Some of those things may not be relevant to std benchmarks, which are mostly are CPU- or memory-bandwidth-bound and single-threaded. They shouldn't suffer much from swap, IRQs or SMT-siblibgs if you ensured the system is mostly idle since they depend on system activity (well, depends on how many cores one has... maybe core isolation is still worth it).
Scheduling and throttling have the biggest impact in my experience. If we had a benchmark that tried to do a parallel sort on a huge dataset that would be a different story.
Adjusting the scaling governor is a good point.
- mention scaling governors - linking stage0 as rustup toolchain is now supported
|
Thank you! |
generated from commit b61d0a2
@jyn514 requested a guide how to benchmark std changes.