The Kubernetes Device Plugin allows Kubernetes to discover and manage NVIDIA GPUs. The NVIDIA Container Runtime is a low-level library that provides the necessary hooks to expose the GPUs to containers. The NVIDIA driver is the foundation for all GPU operations. Kube-proxy mode and containerd CRI are important for general kubernetes networking and containerization but do not specifically enable GPU Support. IPVS is not specifically related and Containerd is not NVIDIA specific

The correct answers are A and E. RBAC allows you to control who can create, modify, or delete GPU resources within each namespace. It defines what actions (verbs) are allowed on what resources. Combined with a custom admission controller to validate the GPU request, you can enforce the access control across the namespaces. Implementing an admission controller enforces policies and validates requests to access GPU resources originated from authorized namespaces. Option B doesn't exist. Option C, although helpful for network segmentation, does not directly control GPU access. Option D only limits usage, not access.

B, C, and E are correct. 'nvidia-smi' provides real-time GPU memory usage. Application logs often contain CUDA errors indicating memory issues. Nsight Systems offers detailed profiling to pinpoint memory leaks. A is not relevant to GPU memory leaks. D is a workaround, not a diagnostic solution.

Driver upgrades can introduce CUDA toolkit incompatibilities or changes in the CUDAAPI that necessitate recompilation or code modification.

Data caching on NVMe drives significantly reduces the need to repeatedly fetch data from object storage. Introducing a distributed filesystem allows a central point where to access the objects. Parquet and Apache Arrow are optimized for columnar data that can be used for parallel access and loading of data from an object store.