With NVIDIA System Profiler it is possible to trace how the virtualization software stack works.
Flash your DRIVE devkit with a hypervisor (option -H
for bootburn.sh
).
(When using DriveInstall, this happens by default.)
Please note that not all virtual configurations support tracing events from
the hypervisor. linux
and linux-linux
are known to work.
At least one of the guest machines needs to be Linux, and there should be mempools (shared memory) mappings between event providers and this Linux guest VM.
Enable hypervisor trace events using the hypervisor console. Please refer to the PDK documentation for further details.
In NVIDIA System Profiler, select Collect HV trace checkbox.
Next, specify the location of the pct.json
file corresponding to the used
virtual configuration.
Run a profiling session as usual.
In the resulting report, expand the Cross-hypervisor trace row to see the events. Detailed information is available in tooltips.
Note: CPU cores use different numbering inside guest VMs and in the hypervisor trace rows:
Hypervisor trace uses physical core numbers.
Guest VMs uses virtual core numbers, which might not map directly to the physical core numbers. On Linux guests, it is customary that the first CPU core number is always 0.
NVIDIA® System Profiler Documentation Rev. 3.9.170817 ©2017. NVIDIA Corporation. All Rights Reserved.