Nvprof query-able metrics don't include any Unified Memory metrics?

I’m trying to find a way to query the number of page faults (and associated unified memory stats) as the kernel is running. Yes, I know that the Visual profiler does this already.

However, I’m attempting to create some time series data plots around my workload’s unified memory performance as it progresses in its execution. I was hoping to create some sort of a CPU based solution that’ll query hardware counters for this sort of information as my workload runs. But, I see no such query-able metrics for unified memory that can be accessed via the Profiler’s Driver API. Is there another way to get unified memory metrics realtime?

http://docs.nvidia.com/cuda/profiler-users-guide/index.html#metrics-reference

Hi, kmankad

I’m afraid there are no query-able metrics in unified memory for now.

The unified memory activities come with timestamp and hence can be correlated with the kernel in future release.