Potential speed regression about pysr HOT 5 CLOSED

Quick update: it seems like adding @inbounds to the following loop:

@inbounds for i=1:npopulations

seems to help significantly. This is one difference with the Threads code - I declare @inbounds at the start of this loop.

It seems as though Julia might have been checking for race conditions otherwise. Yet, I know that race conditions aren't possible in this section of the code, since the hall of fame update is only done by the head worker, so @inbounds seems to help the compiler realize that.

This is essentially a 2-3x speedup!

Also: I need to add a proper benchmark test. I like numpy's tool https://asv.readthedocs.io/en/stable/ since it can be used to measure speed regressions across a commit history.

from pysr.

Seems like there are still significant speed reductions with a recent commit... See the cycles per second of travis build 39: https://travis-ci.com/github/MilesCranmer/PySR/builds/186862848, which gives 5.030e+04 cycles per second.

But the most recent travis build https://travis-ci.com/github/MilesCranmer/PySR/builds/187199749 gets 1.870e+04 cycles per second for what (I think) is the same test...

from pysr.

Very strange... maybe it's a missing @async somewhere?

See diff in julia/sr.jl here: 121e6ac...master

from pysr.

Okay, fixed on master. Will push new pip version soon.

The issue was the nested recursive threading introduced in 3a578d2.

Reading this post: https://julialang.org/blog/2019/07/multithreading/, where they do this same nested recursive threading to compute Fibonnaci numbers, made me think that it could be used for all operations on binary trees in PySR. It is supposed to be fine with millions of threads, and threads calling threads calling threads...

But for whatever reason it seems to have hurt the performance a lot - perhaps it doesn't play nice with Distributed also being used.

from pysr.

Benchmarks on v0.3.X with 4-core runs have been added to the benchmarks/ folder. They are copied below:

So v0.3.15 has gotten rid of the speed regressions throughout v0.3.

from pysr.