Capability | HPC & supercomputing
Clusters for simulation, modelling and science.
A cluster's real performance is decided by the balance between nodes, network and storage — not by the peak FLOPS on the datasheet. We size all three together, against your actual workload, and prove the result through benchmarking and acceptance testing.
What it covers
- Architecture
- Workload-sized CPU / GPU node balance — memory & I/O ratios matched to the job mix
- Interconnect
- InfiniBand / RoCE fat-tree & spine-leaf fabrics, rail-optimised, NCCL-tuned
- Storage
- Parallel file systems — Lustre, Spectrum Scale, BeeGFS — with tiered NVMe
- Provisioning
- Cluster deployment & image/driver control (Warewulf, xCAT, Base Command)
- Scheduling
- Slurm / PBS with fair-share, accounting & per-job telemetry
- Validation
- HPL, HPCG, STREAM & MLPerf benchmarking and acceptance testing
From national-scale practice
Large systems are engineered as fleets, not collections of clusters. The same discipline shaped our engineering dossier for DEUS 11, a national-scale supercomputer programme in Cluj-Napoca.
- Partitioning
- Heterogeneous multi-vendor partitions per workload class — one fabric, one storage system, one scheduler
- Job routing
- Feature-tagged constraints, transparent routing across partitions — the condo model puts idle capacity to work
- Host CPUs
- Chosen per partition role — dense cores for data marshalling, high clocks for serial fractions
- Fleet baseline
- Firmware & software stack unified fleet-wide — spares commonality, a single baseline to maintain
- Expansion
- Floor, power & cooling reserved for next accelerator generations — rack fit-out deferred, no stranded cost
- Data protection
- Erasure coding matched to data classes — long-term archive with WORM & air-gap