City wants several 311 microservices to share A100 GPUs without starving each other. What should admins configure?

Multi-Instance GPU (MIG) profiles partition an A100 into isolated GPU instances with separate memory and compute paths. For City 311 microservices that must share accelerators without starving each other, MIG gives each service a fenced slice with predictable QoS instead of one noisy shared context. Configure MIG profiles that match municipal service sizes, expose the corresponding resource names through the GPU Operator/device plugin, and request those resources in each Deployment.

Pinning every microservice to the same full GPU with no isolation recreates contention: one bursty classifier can inflate latency for every peer. Disabling the GPU Operator so no pods can request accelerators removes the device plugin path entirely—311 cannot schedule GPU work at all. Oversubscribing each full GPU with time-slicing only and no MIG planning may improve utilization for tolerant batch jobs, but time-slicing does not provide the memory/fault isolation MIG is designed for when interactive civic SLAs matter.

Exam caveat: MIG is supported on specific NVIDIA architectures (for example A100/H100 families); not every GPU in the civic fleet can use it. Changing MIG geometry usually requires draining the node and resetting GPU mode. Mixing MIG and full-GPU scheduling in one pool without clear labels confuses the scheduler—standardize per node pool.

Operational check: after enabling MIG on the 311 node category, confirm nvidia-smi -L shows the expected instances, pods request the correct MIG resource, and latency/error budgets hold under concurrent microservice load. Document which municipal services map to which profile sizes for capacity planning.