search

Syllo / nvtop

GPU & Accelerator process monitoring for AMD, Apple, Huawei, Intel, NVIDIA and Qualcomm
Other
8.16k stars 292 forks source link

Add columns for per-process GPU, Encoder, and Decoder usage #91

Closed philipl closed 3 years ago

philipl commented 3 years ago

These stats exist, although they appear to behave strangely. They seem to take a significant amount of time to rise from zero to the correct value, although they do seem to eventually converge.

Syllo commented 3 years ago

Hey,

Nice to see this being investigated by someone. And with a pull request on top of that, thanks!

As you pointed out, the stat doesn't seem to reflect the current state of the GPU, but seems to be lacking behind.

I'll play with it to see if the latency can be reduced. If not I might include it behind a command line option with a warning about the latency.

Syllo commented 3 years ago

Hi,

I found out why the values were lagging behind. The function nvmlDeviceGetProcessUtilization takes as last argument a lastSeenTimeStamp parameter which needs to be stored from a previous call to get the utilization since the last call. Otherwise the function returns the utilization over a larger period of time which depends on the driver.

I did a big overhaul of the internals recently and added the per-process utilization with the fix at the end.

Thanks for pointing this out

Take care