- A
- A
- A
- ABC
- ABC
- ABC
- А
- А
- А
- А
- А
HSE’s Supercomputer Reaches Peak Performance at One PetaFLOPS
© iStock
Here’s more about how the supercomputer’s performance capacity was enlarged, and what new opportunities this brings.
On December 15, HSE’s supercomputer reached a peak performance of 1 petaFLOPS: the cluster was extended by 11 compute nodes based on Intel Xeon Gold 6248R processors and three nodes based on Xeon Gold 6240R processors with graphic accelerators NVIDIA Tesla V100-32GB. PetaFLOPS means quadrillion (thousand trillion) floating point operations per second (FLOPS).
Pavel Kostenetskiy, Head of the Supercomputer Modeling Unit
We have been working on extending the HSE supercomputer since mid-November 2020: we did it carefully, without switching it off, so that researchers and teachers could continue their important work.
The need to extend the supercomputer became obvious when in August 2020 its monthly load had reached its maximum capacity levels. Researchers’ tasks had to wait in a queue. The load has grown due to the growing number of HSE researchers and teaches who are competent in the use of modern technology; it is also a successful outcome of work performed by the Supercomputer Modeling Unit engineers, who maintain stable operations of the high-performance system.
‘Throughout the year, we have been gathering statistics on the use of the supercomputer’s computing nodes by HSE users. The statistics helped us realize that users needed two times more CPUs than the cluster has and only 30% more graphic processors. Based on this conclusion, we developed an optimal extension set,’ Pavel Kostenetskiy said.
The set includes 11 nodes with powerful CPUs and three nodes based on graphic processors, which are similar to what the cluster already has. The equipment has been chosen so that all cluster nodes are compatible and work in one computational field. Users can choose the node type for their tasks and, if necessary, they can perform a bigger task on several nodes of different types simultaneously. CPU-based nodes are sufficient for every other task launched on the HSE cluster. Their addition will help relieve the more expensive graphic nodes from nonspecific tasks, which makes the extension more advantageous.