Hitesh Sahu

The Unix fundamentals underneath every container and Kubernetes pod — the fork/exec/wait process model, file descriptors and the everything-is-a-file philosophy, the syscall boundary between user space and the kernel, signals, pipes, process memory layout, and the permission model.

Slurm from the ground up — the controller/node-daemon architecture, partitions and QOS, job submission with sbatch/srun/salloc, GPU allocation with GRES, multi-node MPI jobs, running containers under Slurm with enroot and pyxis, multifactor job priority, preemption, job arrays and dependencies, node health and cgroup enforcement, topology-aware scheduling, the slurmrestd API, Slurm vs Kubernetes, and a set of interview questions with answers.

How NVIDIA Base Command Manager (BCM) operates an entire GPU cluster — bare metal provisioning and node imaging, the head node vs compute node architecture, Slurm and Kubernetes workload manager integration, user and group management, the cmsh CLI and REST API, and how BCM, DCGM, and SMI fit together at different layers of the stack.

How NVIDIA's Data Center GPU Manager works — the nv-hostengine daemon, key DCGM metric field IDs, XID error codes and diagnostic levels, the DCGM Exporter DaemonSet and Prometheus integration on Kubernetes, PromQL queries and alert rules for GPU clusters, DCGM-driven KEDA autoscaling, and MIG instance monitoring.

CI/CD from the ground up — Continuous Integration vs Delivery vs Deployment, pipeline stages, GitHub Actions and Jenkins pipelines, push-based CD vs GitOps with ArgoCD, a real GPU inference deployment pipeline, and a set of interview questions with answers.

Ansible from the ground up — agentless push architecture, inventory, playbooks, modules, roles, idempotency, Ansible Vault, a real GPU-node provisioning example, Ansible vs Terraform, and a set of interview questions with answers.

The Unix fundamentals underneath every container and Kubernetes pod — the fork/exec/wait process model, file descriptors and the everything-is-a-file philosophy, the syscall boundary between user space and the kernel, signals, pipes, process memory layout, and the permission model.

Slurm from the ground up — the controller/node-daemon architecture, partitions and QOS, job submission with sbatch/srun/salloc, GPU allocation with GRES, multi-node MPI jobs, running containers under Slurm with enroot and pyxis, multifactor job priority, preemption, job arrays and dependencies, node health and cgroup enforcement, topology-aware scheduling, the slurmrestd API, Slurm vs Kubernetes, and a set of interview questions with answers.

How NVIDIA Base Command Manager (BCM) operates an entire GPU cluster — bare metal provisioning and node imaging, the head node vs compute node architecture, Slurm and Kubernetes workload manager integration, user and group management, the cmsh CLI and REST API, and how BCM, DCGM, and SMI fit together at different layers of the stack.

How NVIDIA's Data Center GPU Manager works — the nv-hostengine daemon, key DCGM metric field IDs, XID error codes and diagnostic levels, the DCGM Exporter DaemonSet and Prometheus integration on Kubernetes, PromQL queries and alert rules for GPU clusters, DCGM-driven KEDA autoscaling, and MIG instance monitoring.

CI/CD from the ground up — Continuous Integration vs Delivery vs Deployment, pipeline stages, GitHub Actions and Jenkins pipelines, push-based CD vs GitOps with ArgoCD, a real GPU inference deployment pipeline, and a set of interview questions with answers.

Ansible from the ground up — agentless push architecture, inventory, playbooks, modules, roles, idempotency, Ansible Vault, a real GPU-node provisioning example, Ansible vs Terraform, and a set of interview questions with answers.