Build Python polars wheels with PGO

[messense]

Problem description

It seems that PGO is a success for pydantic-core, I wonder whether it will speedup py-polars.

[ritchie46]

Yes, I'd be very interested in that. I think including the db-benchmark and the tpch benchmark code in the guided data seem to me to be interesting candidates.

Tests would be easiest. Not any experience with setting it up though, so any help on this is appreciated.

[messense]

@ritchie46 Is there any easy way to compare two benchmark results? I'd like to confirm that it actually improves performance before putting too much effort.

[ritchie46]

The easiest start are the TPCH benchmarks: https://github.com/pola-rs/tpch/tree/main/polars_queries

The repo has a Makefile that creates the dataset. Running this provides a timing table and some plot utils. Now that I think of it, that's also good PGO input.

[zamazan4ik]

I think you also could be interested in my recent benchmark regarding PGO to different kinds of software (including a lot of databases and database libraries (like RocksDB)) - https://github.com/zamazan4ik/awesome-pgo .

[jmakov]

From some of the blog posts it looks like +15% can be done. That's the difference a CPU makes from one generation to another.

[FilipAndersson245]

Any update on this? PGO is a very interesting way to improve performance with the only cost is deployment and build complexity.
I can recommend looking at https://github.com/Kobzol/cargo-pgo for inspiration on how to run it.

[alonme]

I started taking a look at that today.

It seemed that using the flow documented in https://doc.rust-lang.org/rustc/profile-guided-optimization.html - and when benchmarking the PGO'd version after that - i didn't any change in performance.

I suspect that the issue is that only the py-polars crate was actually compiled to generate the profiling data and then optimized.

Currently trying to make it work on all the crates

(Note - I use rosetta2, However i don't think that should be a problem or a reason for the PGO not to make an affect)

[FilipAndersson245]

I have only used Kobzols Cargo-pgo tool so i have less experience with the raw pgo scripts. @zamazan4ik the go to guy with everything pgo.

[alonme]

After Compiling with global flags e.g. export RUSTFLAGS="-Cprofile-generate=/tmp/pgo-data and then export RUSTFLAGS=-Cprofile-use=/tmp/pgo-data/merged.profdata errors when compiling the second one with

!"CG Profile"
rustc-LLVM ERROR: Broken module found, compilation aborted!
error: could not compile `py-polars` (lib)

which is referenced in the following issues:
rust-lang/rust#115344
rust-lang/rust#117220

I will try to check if a different LTO setting will make it compile, and check the effect

[alonme]

lto = "thin" compiles successfully 👍

[alonme]

After Compiling with global flags e.g. export RUSTFLAGS="-Cprofile-generate=/tmp/pgo-data and then export RUSTFLAGS=-Cprofile-use=/tmp/pgo-data/merged.profdata errors when compiling the second one with

!"CG Profile"
rustc-LLVM ERROR: Broken module found, compilation aborted!
error: could not compile `py-polars` (lib)

which is referenced in the following issues: rust-lang/rust#115344 rust-lang/rust#117220

I will try to check if a different LTO setting will make it compile, and check the effect

these rust issues have been resolved - would probably be a good idea to try again

[FilipAndersson245]

@alonme did you manage to build with lot and pgo?, I know you looked into this earlier.