Tools

Interactive utilities for triage, validation, and authoring. Inspired by an internal support-tooling site, ported and generalized.

GPU node health

Multi-check on a GPU node: nvidia-smi, NVLink, ECC, PCIe link speed/width, fan/thermal, persistence mode, fabric-manager status.

Weka mount checker

Validate a Weka mount: cluster reachability, drivers-loader, wekafsio refcount, /proc/mounts entries, common stale-mount fixes.

ulimit analyzer

Walk through /etc/security/limits.conf, systemd unit overrides, and the effective ulimits for kubelet, containerd, slurmd, and user processes.

Pod inspector

Cheat sheet for diagnosing a stuck pod: events, scheduler reasons, init containers, volume mount errors, scheduler webhook denials.

Slurm job debugger

Step-by-step recipe to debug a stuck or failed Slurm job: scontrol show, NodeFail/NodeNotResponding, prolog/epilog hooks, enroot/pyxis errors.

Node speed benchmark

Standard mini-benchmark suite for a fresh GPU node: nccl-tests all-reduce, ib_send_bw / ib_read_bw / ib_write_bw across rails, fio against local NVMe and shared FS.

kubectl helper

Common kubectl one-liners: drain, cordon, taint, debug ephemeral container, evict, exec into init container, port-forward, cp, etc.

fio VAST tool

fio recipes tuned for VAST: throughput, IOPS, mixed workloads, multi-job, target queue depth, AIO vs libaio vs io_uring.

Cross-DC rsync

rsync with multi-stream, bandwidth limits, partial-resume, checksum-mode, sshpass alternatives, performance profiling for inter-datacenter copies.

Grafana link generator

Build deep links to Grafana dashboards with time range, variables, and node selectors prefilled — useful for paging on-call to the right view.

Slurm commands

sinfo / squeue / scontrol / scancel / sacct cheat sheet with the most useful flag combinations.

Slurm job submission

Templates for sbatch scripts: GPU jobs, multi-node MPI, enroot+pyxis, Slurm + container runtime, common pitfalls.

Speed test commands

curl, iperf3, ib_send_bw, perftest one-liners for measuring network performance from inside a pod or on the bare metal node.

Tailscale

Tailscale on bare metal hosts: subnet routes, exit nodes, ACLs, MagicDNS, troubleshooting.

Markdown viewer

Paste a markdown blob and render it. Useful when ops runbooks live in private repos / Slack and you want a quick clean view.

Prompt library

Curated prompts for LLMs to assist with operational tasks: incident summary, runbook drafts, diff explanation, postmortem structure.