Needhelp with app_config and GPU usage
14 May 2023 2:04:33 UTC
Topic 229514
(moderation:
So I run Einstein and LHC at the same time. LHC only has CPU work and I only run Einstein on GPU. When I run them together it often puts LHC on all 12 threads and has the GPU share with LHC. I specified 0.5 GPU but it will only run one at a time unless I suspend LHC. The Einstein GPU units run really slow this way, I want to run at least one and try two units at a time with their own CPU thread, no sharing.
If I do this:
<app_config>
<app>
<name>einstein_O3MDF</name>
<max_concurrent>2</max_concurrent>
<report_results_immediately/>
<fraction_done_exact/>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
<report_results_immediately/>
</app_config>
It "claims" to running 1 CPU per 0.5 GPU but it isn't, it is still sharing and mostly runs one instance.
I would be thankful for any help.
You need to put a
)
You need to put a max_concurrent on the LHC tasks so they don't hog all the cpu and leave a couple of cpu threads for the two gpu tasks.
I'll try that, thank you. As
)
I'll try that, thank you. As a follow up it seems the Atlas simulations are the ones that step over the line.
In LHC the Atlas tasks are
)
In LHC the Atlas tasks are the only multi core tasks. You can specify the number of cores used by the task on LHC project preferences. The number of simultaneous Atlas tasks you can specify in app_config.xml as Keith Myers said.