The 'nodeselector' is used to target the deployment to nodes with label 'accelerator: nvidia-tesla-t4'. "nodeAffinity' is a more advanced way of doing this and is recommended, but in the absence of explicit nodeAffinity, nodeSelector is sufficient.

Increasing network bandwidth ensures data can be transferred faster. Migrating to a faster storage tier improves I/O performance directly. Optimizing data pipelines reduces the overall amount of data that needs to be read/written. Reducing GPUs will not solve I/O bottleneck, just slow down training. Aggressive compression can add overhead.

A multi-node cluster provides redundancy and failover capabilities, ensuring high availability. Geographically diverse edge deployments minimize the impact of regional outages. A single server (A) is a single point of failure. Backups (C) are important but don't prevent downtime. Monitoring (D) helps identify issues but doesn't ensure resilience. Solely relying on edge devices (E) limits manageability and control.

You are running a Docker container that utilizes NVIDIA GPUs for deep learning inference. The application inside the container fails to detect the GPUs. You've verified that NVIDIA drivers are installed on the host. What is the MOST likely cause and how do you fix it?

The NVIDIA Data Center GPU Manager (DCGM) exporter is specifically designed to collect and expose GPU metrics in a format that Prometheus can consume. If GPU metrics are not being collected, the DCGM exporter is the most likely culprit. The other options are less directly related to GPU metric collection. Option A pertains more to core Kubernetes metrics, option C relates to generic prometheus service discovery which isn't specialized to GPU data. Logging drivers and API throttling are less likely to directly block metrics collection.