Compile with PGO + Bolt

[zamazan4ik]

Who is this for and what problem do they have today?

For the end-users, who want to use even more performant Redpanda installations.

What are the success criteria?

• Profile-Guided Optimization (PGO) and LLVM BOLT are integrated into the build scripts
• Prebuilt Redpanda binaries are built with PGO and Bolt

Why is solving this problem impactful?

Could be useful for Redpanda performance.

Additional notes

Not sure about the current build defaults but enabling LTO (preferably "fat" version) would be beneficial too. Much more results about enabling PGO in different projects can be found here.

JIRA Link: CORE-1126

[jcsp]

I'm interested in whether you're suggesting this as an experiment, or if you tried it and found a measurable improvement?

[zamazan4ik]

Now I am suggesting it as an experiment. I have applied PGO multiple times to different software (like Clang, Vector, etc) with good results in performance, but didn't try yet to do it with Redpanda.

[zamazan4ik]

Oh, and I completely forgot to suggest trying to optimize compilers on CI for faster build times: https://github.com/scylladb/scylladb/issues/10985 . I have done the same trick in other projects with the same results.

[emaxerrno]

@zamazan4ik - probably using better CPU profiles than what we are using would give you the biggest bang for the buck.

[zamazan4ik]

I think these both ways can be used at the same time. Since my tests show pretty good performance improvements from enabling PGO for other software, I am still interested in optimizing Redpanda with PGO. However, the more ways to speed-up Redpanda - the better.