I have held roles as HPC architect, engineer, group lead and IT director, in environments where the machines are expensive and the deadlines are external. The work runs from architecture and procurement through build-out, benchmarking and the operational life of a cluster afterwards.

What I work on

• Compute and provisioning — hardware selection, node classes, image and provisioning lifecycle, capacity planning.
• Fabric — InfiniBand from EDR through NDR, fat tree and dragonfly topologies, OpenSM routing engines, RoCE and RDMA, switch and adapter firmware.
• Parallel storage — GPFS and IBM Storage Scale, Lustre, WekaFS, VAST, Isilon, DDN, BeeGFS and Ceph. Policy and tiering, data movement, and benchmarking that measures the filesystem rather than a client cache.
• GPU and AI stacks — A100 and H100 systems, NVLink, DCGM, CUDA and ROCm, NCCL collective tuning, containerised delivery.
• Schedulers — Slurm, PBS Pro, LSF and Grid Engine. Partition and QoS design, GPU and licence-aware scheduling, fair share and accounting.
• The enterprise IT around it — identity, monitoring, patching and the compliance the cluster has to live inside.

Writing

I write up the failure modes that cost time, with the command that distinguishes each one. Most articles exist because something looked fine and was not: a benchmark measuring a cache, a counter read as a rate, a check that was green because it was pointed slightly to the side of the fault.

All 72 articles · Browse by topic · RSS

Contact

Assessment and engineering work is performed under written authorisation. Get in touch.