HMM Breaks Process Memory Metrics reporting #106
Description
Is this a known issue?
I have a sample CUDA workload that is allocating memory.
Before that starts I setup nv-hostengine to run as root and then start StartPIDWatch.
In another process running within a container, I am then running that sample CUDA workload and it is collecting its own process statistics with GetProcessInfo.
If I run this on A10 and T4 GPUs with the 580 series driver, I get GlobalUsed memory metrics.
When I run this on an L4 GPU with HMM addressing mode enabled, GlobalUsed always reports 0. Is this a known deficiency? With an updated driver, kernel, etc. on T4 and A10 GPUs, HMM isn't enabled and we don't have these issues.