A WARN log should be added at least (logged once on startup) that notifies the user about a potentially offending configuration: GPU isolation is enabled but LCE is disabled.
I think this is a dangerous, yet valid configuration: As LCE is the only container executor that utilizes cgroups, no real HW-isolation happens if LCE is disabled.
Let's suppose we have 2 GPU devices in 1 node:
- NM reports 2 devices (as a Resource) to RM
- RM assigns GPU#1 to container#2 that requests 1 GPU device
- When container#2 is also requesting 1 GPU device, RM is going to assign either GPU#1 or GPU#2, so there's no guarantee that GPU#2 will be assigned. If GPU#1 is assigned to a second container, nasty things could happen.