Evaluate using Profile-Guided Optimization (PGO) and Post-Link Optimization (PLO) for SteamOS

[zamazan4ik]

This issue is not exactly about the problem - it's just an idea of how to improve SteamOS performance and UX and battery life for SteamDeck.

I can imagine multiple things where PGO can shine in SteamOS:

• The Linux kernel itself since it's a core component for SteamOS:
  • Paper
  • Microsoft presentation
  • ASOS (Application Specific Operating System)
  • TCP Stream perf
  • Phoronix post
  • An attempt to PGO Linux kernel: http://coolypf.com/kpgo.htm
  • Gentoo Wiki
  • Optimizing Linux kernel with Clang. An article(in Russian) and results
• Things like shader compilers could benefit from PGO. Having faster shader compilation could improve first-user UX and battery life. Some benchmarks are done for Wgpu. An issue about PGO in glslang.
• Optimize with PGO Mesa drivers and tooling: Phoronix post
• Maybe Proton could benefit from PGO but I have no benchmark results for it right now so I'm just guessing here.

Maybe they are valuable for SteamOS applications. Here I collect as many PGO cases as I can find - you can check them and find more other applications.

[kisak-valve]

Hello @zamazan4ik, this is a recurring topic among many open source projects. The general outcome is that you can make one very specific use case marginally better while harming almost all other use cases and sometimes creates some extremely poor performance results when the profile used to compile doesn't match reality.

When generalized to an entire OS install, this is a sink hole to burn developer time and isn't worth looking into. The time spent tuning profiles is better spent integrating newer package versions which have picked up optimizations as part of their upstream maintenance. If there's a specific and significant performance improvement to be found in a specific project, then contribute it upstream for everybody to benefit.

[zamazan4ik]

marginally better

I have a lot of examples where the outcome is significant, not marginal. All the benchmarks you can check here (the same link was in the first message).

while harming almost all other use cases

It's not true if you are using several profiles from different workloads and use all of them simultaneously to create one PGO build (it could be done via merging several PGO profiles into one via dedicated tooling like llvm-profdata merge).

sometimes creates some extremely poor performance results when the profile used to compile doesn't match reality.

I do agree. However, it is usually simply mitigated by adding this profile to the PGO training set as well.

When generalized to an entire OS install, this is a sink hole to burn developer time and isn't worth looking into. The time spent tuning profiles is better spent integrating newer package versions which have picked up optimizations as part of their upstream maintenance. If there's a specific and significant performance improvement to be found in a specific project, then contribute it upstream for everybody to benefit.

Yeah, makes sense. Thanks a lot for the answer!