Performance Regression when using Threads Backend

[alexishuante]

Hello,

It seems that currently there is an issue with JACC v0.1.0 where the performance when using the Threads Backend is worse compared to JACC v0.0.5.

When running a kernel, I got an average iteration time of 5.77 seconds when using JACC v0.0.5. Whereas for JACC v0.1.0, I got an average iteration time of 546 seconds.

When I ran a profiler, it seemed pointing to a possible task management overhead, meaning the threading backend is spending a lot of time scheduling and managing tasks.

Let me know if there is anything I can do to help.

Thank you.

[williamfgc]

Hi @alexishuante , please provide the kernel code for reproducibility. Thanks!

[alexishuante]

Here is the kernel code I used: https://github.com/jhonnygzz/JACC_APPS/tree/main/miniBUDEJACC
To reproduce these results, the instructions are provided in the README.md. I also used export JULIA_EXCLUSIVE=1 and export JULIA_NUM_THREADS=128.

Thank you.

[williamfgc]

@alexishuante please feel free to test with the latest main branch and close this issue if fixed. Thanks!

[alexishuante]

I can confirm that JACC v0.1.1 fixes the regression problem in threads. I made a mistake in my run test case. I will close this issue since the problem is now fixed, thank you.