use half of the available cores for hipBLASLt - #281
Conversation
|
Let's give it a try bot: build repo:eessi.io-2025.06-software instance:eessi-bot-aws-eu-south on:arch=x86_64/amd/zen5 for:arch=x86_64/amd/zen5,accel=amd/gfx908+amd/gfx90a+amd/gfx942+amd/gfx1030+amd/gfx1100+amd/gfx1101+amd/gfx1200+amd/gfx1201 |
|
New job on instance
|
|
Not sure what's going on with these jobs. It doesn't look like they run out of memory, but it's getting stuck at some point, and ultimately Slurm aborts and requeues the job on the same node? On the node itself I only see these messages in the syslog: |
|
bot: build repo:eessi.io-2025.06-software instance:eessi-bot-mc-aws on:arch=zen2 for:arch=x86_64/amd/zen2,accel=amd/gfx942 |
|
New job on instance
|
|
bot:cancel jobid:186051 |
|
It did use bot: build repo:eessi.io-2025.06-software instance:eessi-bot-mc-aws on:arch=zen2 for:arch=x86_64/amd/zen2,accel=amd/gfx942 |
|
New job on instance
|
|
bot:cancel jobid:186052 |
|
Still OOM killed, there's a main process taking up a lot of memory, and 4 joblib workers/threads. This was just before the main process got killed: |
|
Related issue: ROCm/hipBLASLt#1614 |
|
Not sure how to deal with this, let's give it one more try with only 2 cores. |
|
bot: build repo:eessi.io-2025.06-software instance:eessi-bot-mc-aws on:arch=zen2 for:arch=x86_64/amd/zen2,accel=amd/gfx942 |
|
New job on instance
|
|
Now the job itself failed with a bot: build repo:eessi.io-2025.06-software instance:eessi-bot-mc-aws on:arch=zen2 for:arch=x86_64/amd/zen2,accel=amd/gfx942 |
|
New job on instance
|
|
Trying with bot: build repo:eessi.io-2025.06-software instance:eessi-bot-mc-aws on:arch=zen2 for:arch=x86_64/amd/zen2,accel=amd/gfx942 |
|
New job on instance
|
|
Just noticed that the easyconfig already includes that option, so this is probably not going to change anything... |
|
Also found ROCm/rocm-libraries#316 and easybuilders/easybuild-easyconfigs#23542, where similar issues are discussed. Building for |
|
I've started an Besides from the increased memory usage, note that it also took much longer compared to the successful builds for other targets in EESSI/software-layer#1549. |
|
Closing this, as it's not going to fix the issue. Summarized the observations from this PR in EESSI/software-layer#1549 (comment). |
This should hopefully prevent it from running out of memory.