- This topic has 2 replies, 3 voices, and was last updated 1 day, 22 hours ago by
.
-
Posts
-
Hello,
I am running distributed GPU training experiments (PyTorch FSDP / NCCL all_reduce) on a FABRIC slice with two NVIDIA A30 GPUs at the PRIN site. Our main task is to compare Fabric training times with with those from our own simulator.
For reproducible bandwidth measurements I need to eliminate CPU noise from hypervisor co-scheduling. I understand that GPU components use PCIe passthrough and are therefore exclusive to my slice. My question is:
Is there a way to request a node where the physical CPU cores are also exclusively assigned to my VM (no hypervisor overcommit / no co-tenancy)? I did not find a dedicated_cpu or exclusive parameter in the FABlib API.
Site: PRIN
Node type: VM with 2× GPU_A30Thank you.
There is a feature in fablib that allows for cpu/core pinining. That is as close as you can get to core exclusivity i guess.
You can try the cpupin_common script that eliminates four of five layers of CPU interference:
The NIST-MQNS artifact contains a concrete example of using this script to perform a CPU-bound benchmark of a Python application.
- You must be logged in to reply to this topic.