FLUTE GPU utilisation vs performance

[microsoft]

FLUTE GPU utilisation vs performance

[simra]

Glad to hear you are running some benchmarks- we'll look forward to hearing about your results.

Over the fall we invested a lot of effort into improving the GPU utilization, and we saw about 30% performance gains in the per-round duration. However we haven't recently investigated any outstanding bottlenecks or how to address them.

One thing that might provide some clues is to enable profiling in the server or client. Both server and client configurations have a do_profiling option which will enable the python profiler and also compute some durations of interest and log them as metrics. I would start with only enabling this in the server since profiling in the client can become very chatty. Let us know if you need any more details and we can walk you through through setting up the config and interpreting the results.

[Iacob-Alexandru-Andrei]

Thank you very much for your quick reply, I will make sure to contact you if we need further information following the profiling.