• Production system
• In service
• Discipline-specific system for High-Energy Physics
• Funded by STFC, UKRI, DSIT
• Partitions
• 114 nodes, with 4 NVIDIA A100 40GB accelerators per node
• Manufactured by Atos
• Scheduler: Slurm
• 65 nodes, with 4 NVIDIA A100 80GB accelerators per node
• Manufactured by Atos
• Scheduler: Slurm
• Interconnects: Infiniband HDR, NVLink

Tursa is a GPU cluster hosted at EPCC on behalf of DiRAC. Comprising 181 x Atos BullSequana nodes, 114 with 4 x NVIDIA A100-40GB GPUs and 65 with 4 x NVIDIA A100-80 GPUs, it was designed for performing strong scaling computations in lattice quantum field theory.

GPUs are linked by NVLink within each node, and each GPU has a dedicated fabric adapter with GPUdirect RDMA, allowing for very good strong scaling performance.. The machine has also been demonstrated to perform well for machine learning applications, including training of large language models.

Documentation

• https://epcced.github.io/dirac-docs/tursa-user-guide/
• https://www.dirac.ac.uk

Gaining access

Access for production workloads is via an annual call by the DiRAC Resource Allocation Committee. Seedcorn and benchmarking access is available on an ad-hoc basis via “Director’s discretionary allocations”, requested via email to the DiRAC director.

Restrictions

Nodes must be allocated in powers of two. Jobs may run for up to 48 hours while budget is available. When budget is exhausted, low-priority access is available for up to four queued jobs at a time, with a wall-time limit of 24 hours.