GPU jobs perform poorly on a two-socket server. Which setting do you check?
Why C is the answer
NUMA and CPU affinity
Keep CPU threads near their GPU and NIC.
In detail
- Keep CPU threads near their GPU and NIC.
- Keep threads near their GPU and NIC.
Use case
An illustrative scenario, not a real incident.
Jobs run slowly, because threads run on the far CPU socket from their GPU. Fixing affinity helps.
See it in different ways
stateDiagram-v2 [*] --> Symptom Symptom --> Logs: Xid and BMC events Logs --> Diagnostics Diagnostics --> Replace Replace --> Retest Retest --> [*]
mindmap
root(("Two-socket server"))
n1["Socket 0"]
n2["GPU 0"]
n3["NIC 0"]
n4["Its threads"]
n5["Socket 1"]
n6["GPU 1"]
n7["NIC 1"]
n8["Its threads"]Key point: Check NUMA and CPU affinity.
Why the other options are wrong
- A. The GPU driver's release date
- That does not affect placement.
- B. The fan speed profile
- That does not affect placement.
- D. The BMC network settings
- Those do not affect placement.