The University of Melbourne has launched a new High-performance computing (HPC) service called Spartan, combining traditional HPC with a flexible cloud computing component. No other university has put a system like this into production.
Mr Bernard Meade, Head of Research Computer Services at the University of Melbourne, said that the application of HPC and cloud techniques would increase research productivity across a wide range of disciplines.
"Many research projects demand high speed interconnect. Spartan can quickly scale into cloud based virtual machines as needed, and expand the HPC system as user needs evolve," Mr Meade said.
"Traditional HPC systems are typically tailored for a few specific use cases, but in practice are used for a much wider variety of applications, resulting in less than optimal usage," he said.
Cloud systems allow for the sharing of computer resources without the need to be prescriptive as to how they are used. They also allow for the rapid deployment and reclaiming of resources for the shared pool. Spartan is designed to take advantage of the best of both these worlds.
"High level computing has become an integral part of much of the research activity undertaken at the University. Spartan will help us provide a world-class research environment," said Professor Margaret Sheil, Acting Vice-Chancellor at the University of Melbourne.
Spartan can grow and evolve according to the demands of researchers, expanding physically or virtually as required. The unique design features a nucleus of high performance, tightly coupled machines, augmented by thousands of compute cores in the Melbourne Node of the Research Cloud.
Such flexibility offers researchers greater options. They will be able to run jobs on bare metal with high-speed interconnections, use a cluster of virtual machines, or use a combination of the two. The system also caters for the diverse workloads of modern research, rather than forcing them all together in a sub-optimal environment.
Spartan uses the open-source Linux operating system, common throughout the HPC and Cloud industry, and SLURM as a workload manager. Physical hardware includes Intel systems sourced from Dell and switches from Mellanox and Cisco.
The combination of the architecture, hardware, and software choices has already generated impressive results with the fastest latency tests down to 1.15 µsec.
"Spartan is a hybrid HPC service, designed to suit a wide range of problem domains. We believe this is the future of HPC," Mr. Meade said.