Test Profile-Guided Optimization (PGO)

[zamazan4ik]

Is your feature request related to a problem? Please describe. Not a problem - just an idea of how to (possibly) improve ManticoreSearch performance.

Describe the solution you'd like Recently I did a lot of testing Profile-Guided Optimization (PGO) on different kinds of software (including databases). Results you can check here.

I think the same tests should be performed on ManticoreSearch as well since it can improve the performance of the solution. Would be interesting to see the results.

Describe alternatives you've considered Leave things as is and do nothing :)

[tomatolog]

Could you use Docker container to build PGO build first then use PGO result to create optimized build?

As PGO has the meaning for the particular usage pattern and in case daemon load is the same. That means the user should create optimized build that suits only his needs.

[zamazan4ik]

Could you use Docker container to build PGO build first then use PGO result to create optimized build?

Yes, it's possible to use Docker'ized toolchain for doing PGO (via Instrumentation mode).

As PGO has the meaning for the particular usage pattern and in case daemon load is the same. That means the user should create optimized build that suits only his needs.

In general - yes. However, in some cases, it's possible to prepare a "generic enough" workload and use it to optimize delivered to the users' binaries. E.g. in this way, PGO works in Chrome, Firefox, CPython, Rustc.

[tomatolog]

I think there are different usage scenarios:

• plain mode vs RT mode
• plain indexes \ RT indexes \ distributed only node
• PQ only queries \ indexes
• CALL KEYWORDS \ HIGHLIGHT
• replication enabled or disabled
• search full text or fillscan \ select list attributes vs fetch of dockstore \ use of select list expressions
• FACET vs group by vs multi-query
• columnar storage vs row storage
• use of secondary index vs columnar optimization

not sure that all these could be improved in generic case as even switch from GCC to Clang build cause change of performance.

[zamazan4ik]

At least you can try to collect a profile from all these modes, then merge the profiles (via llvm-profdata in the case of Clang), use it during the compilation, and measure the performance compared to the "vanilla" build.

Another option (from which I personally recommend starting) is trying to measure PGO benefits in some scenarios. And if PGO shows something interesting - put a note about PGO into the documentation somewhere in the section "Performance tuning". So ManticoreSearch users will know one more way, how to optimize performance if they are ready to recompile ManticoreSearch.

[tomatolog]

I think this is the excellent task for the community to collect PGO statistics for the section about "Performance tuning"