Spartan is a new breed of HPC; not a cluster but service that manages underlying clusters. These clusters include high performance bare-metal as well as generic cloud based VMs, and many options in between.
It is designed to reduce the time to output for the research community, by providing the most appropriate environment for each specific workload.
Because Spartan has virtualized management and login nodes, it can easily migrate hosts as new hardware is deployed, with partitions expanding according to actual demand, rather than anticipated usage. The totally modular approach eliminates the need for forklift upgrades, as Spartan the service will long outlive the original hardware that supports it.
This presentation will cover:
- The vision for HPC at UoM
- Matching of use cases to compute
- Hardware selection
- Networking and low latency
- Storage
- Identity management
- SLURM, LMod, and Easybuild
- Performance
- Training
- How to combine HPC and Cloud on the same infrastructure
- How to have a flexible modular architecture
- How to profile users and develop a service from their usage patterns
- Using SLURM as a scheduler for cloud instances and cloud burst