Can I possibly query events and metrics of another context with the new cuda toolkit?
I have a long running application and I want to transparently monitor its events and metrics periodically. NVPROF, as useful as it is, does not work in this scenario since its not transparent and requires the application to be run inside NVPROF. I am speaking from sysadmin perspective, I eventually want to use this at the hypervisor level to profile each VMs GPU usage.
I am able to query performance counters of another process on CPU using PAPI. The same approach does not seem to work for the GPU. Any help would be appreciated. Thanks!