Algorithm comparision

[nguu0123]

With different algorithms that we come up with, we will test them with all applications and optasks that we have.
Currently, we have:

• MHD: data movement, data reduction, data movement + data reduction

[rantahar]

• Traditional simulation, computation heavy
  • MHD, GPU heavy
  • other, CPU heavy
• Machine learning training
• Data analysis, data heavy
  • Genetics example?
• Bursty: changing between high and low load

Different types of task sets:

• small number of large tasks, careful scheduling
• large number of small tasks, need to be efficient

[rantahar]

For an AI training application

• https://github.com/Lumi-supercomputer/Getting_Started_with_AI_workshop/tree/ai-202405291

[rantahar]

We have a machine learning benchmark. Runtimes are from a single node on Lumi, which already has 8 GPUs.