Evaluate Profile-Guided Optimization (PGO) and LLVM BOLT

[zamazan4ik]

Introduction
Recently I checked Profile-Guided Optimization (PGO) improvements on multiple projects. The results are here. Since PGO helps with performance for a lot of databases, I think applying PGO to Nebula would be a good idea.

Contents
We need to perform PGO benchmarks on Nebula. And if it shows improvements - add a note about possible improvements Nebula's performance with PGO. Providing an easier way (e.g. a build option) to build scripts with PGO can be useful for the maintainers and end-users too.

Related work
As an additional option, after PGO you can try to apply LLVM BOLT to optimize Nebula even further.

[wey-gu]

Introduction

Recently I checked Profile-Guided Optimization (PGO) improvements on multiple projects. The results are here. Since PGO helps with performance for a lot of databases, I think applying PGO to Nebula would be a good idea.

Contents

We need to perform PGO benchmarks on Nebula. And if it shows improvements - add a note about possible improvements Nebula's performance with PGO. Providing an easier way (e.g. a build option) to build scripts with PGO can be useful for the maintainers and end-users too.

Related work

As an additional option, after PGO you can try to apply LLVM BOLT to optimize Nebula even further.

Dear @zamazan4ik

Welcome to the NebulaGraph community! Thanks a lot for considering boosting NebulaGraph with PGO!

@lipzhu had contributed autoFDO via #4636 and https://docs.nebula-graph.io/3.5.0/8.service-tuning/enable_autofdo_for_nebulagraph/

Could you check if this could be reused/ referred?

Let us know for anything we could collaborate/help!

Thanks!

[zamazan4ik]

Could you check if this could be reused/ referred?

Thanks for the link! Actually, AutoFDO is one of the approaches for doing PGO (AutoFDO is a tool, but the approach with AutoFDO is called Sampling PGO).

I can suggest trying to evaluate another option - Instrumentation PGO. For Clang it's described here. In general, Instrumentation PGO achieves better performance results than Sampling PGO.

As an additional step, I can suggest you try to integrate LLVM BOLT into the build pipeline. LLVM BOLT is able to optimize even after Sampling/Instrumentation PGO.

[wey-gu]

Sorry for the late response, @dutor Instrumentation PGO is quite promising! Shall we do some evaluation manually?