Sampling PGO support

[zamazan4ik]

Hi!

According to the GraalVM documentation, GraalVM supports PGO only with Instrumentation. However, other AoT compilers like Clang, GCC, Rustc support an additional PGO kind - Sampling PGO (also called AutoFDO).

Is there a way to use anyhow sampling PGO with GraalVM? If not, are there plans for implementing sampling PGO in GraalVM?

Having sampling PGO support is important for many PGO-in-production scenarios, especially if we are talking about using PGO at large scales. More about this topic you can find in Google's paper: https://research.google/pubs/pub45290/

Links:

• Sampling PGO in Clang (the same applies to Rustc since Clang and Rustc are both LLVM-based): docs
• Sampling PGO in GCC: GCC Wiki

[axel22]

Yes, the PGO in GraalVM EE does both instrumentation and sampling in the last two releases.
The call-stack samples are included in the .iprof files next to the instrumentation profiles.
On nightlies (didn't check if it made it to the last release), there is an option -H:+BuildReport -H:+BuildReportSamplerFlamegraph that will render these samples inside the build report file (so that they can be inspected visually in a flame graph).
The samples are used to improve some of the EE optimizations.

We will improve our official documentation to explain this better.

[zamazan4ik]

Is it possible in the current sampling PGO implementation to tweak the sampling frequency (like it could be done with perf in the AutoFDO case)?

It's important for us since we can tune the "Sampling overhead <-> Profile precision" tradeoff.

Another question - according to your tests, are there differences between instrumentation and sampling PGO from the optimization opportunity perspective? Because Google reports AutoFDO (sampling PGO) optimization is a bit worse compared to the instrumentation PGO. I am wondering how true it is for the GraalVM PGO implementation.

[zamazan4ik]

@axel22 friendly ping :)