Evaluate Profile-Guided Optimization (PGO) and LLVM BOLT

[kwai]

Evaluate Profile-Guided Optimization (PGO) and LLVM BOLT

[richox]

it sounds attractive, however it's not easy to build a binary for profiling because the native code is built into a lib and dynamically loaded in spark through JNI. we also tried some profiling ways like flamegraphing which does not need a binary, i doubt whether pgo can work like that?

[zamazan4ik]

it sounds attractive, however it's not easy to build a binary for profiling because the native code is built into a lib and dynamically loaded in spark through JNI

You can try to build a "wrapping" binary only for the native part (without JNI stuff), run it on a sample workload, collect the profiles, and then use these profiles during the normal Blaze compilation (PGO-optimize native part + build JNI stuff around it).

we also tried some profiling ways like flamegraphing which does not need a binary, i doubt whether pgo can work like that?

In theory yes, it's possible to do this via Sampling PGO (Clang docs). The same instruction should be valid for Rustc as well but I didn't test such scenario yet - most of my experience is with instrumentation PGO.

UPD: According to the Clang documentation, it's also possible to do instrumentation PGO with shared libraries - the PGO profiles will be dumped for each library. The Pydantic-core project uses this approach for building the PGO-optimized library version.