Profile-Guided Optimization (PGO) and Post-Link Optimization (PLO) for Google Cloud Products

[zamazan4ik]

Hi!

I research a topic about using runtime-based compiler optimization techniques like Profile-Guided Optimization (PGO) and Post-Link Optimization (PLO) across different projects. All my current findings are available in my repo on GitHub: https://github.com/zamazan4ik/awesome-pgo (and a semi-finished article is available at https://github.com/zamazan4ik/awesome-pgo/blob/main/article/article.md). According to my findings, PGO helps with achieving better performance in many cases, including databases (like PostgreSQL, MariaDB, Redis, ClickHouse, SQLite, and many others), log solutions, compilers, browsers, etc.

AFAIK, Google is the largest PGO user (and contributor) in the world (Google uses the AutoFDO approach described at https://research.google/pubs/autofdo-automatic-feedback-directed-optimization-for-warehouse-scale-ap.... Regarding PLO, Google invests resources into Propeller (https://github.com/google/llvm-propeller).

Since GCP provides many open-source projects like databases in a SaaS way, I wonder - does Google perform PGO/PLO optimizations on them internally? I tried to find any information about that but found nothing.

If Google already uses PGO for its open-source-based projects internally - could you please share the performance win numbers from enabling PGO/PLO for these projects? If not, did you consider enabling PGO/PLO for them?

Thank you.

P.S. Not sure if the "Databases" location is the best location for such questions. Feel free to change it if my choice was wrong here.