Simulation of Utilization and Energy Saving of the Angara Interconnect

Lobachevskii Journal of Mathematics. 2022. Vol. 43. P. 873–881.

Mukosey A., Semenov A.

In this paper we address the problem of node allocation for high performance computer systems based on the Angara interconnect with the torus topology. Most allocation strategies for the torus topologies assume redundancy, i.e. for a user job it is possible to allocate more nodes than required. We propose the new node allocation algorithm for supercomputers with the Angara interconnect. The new algorithm removes the limitations of the previously proposed base algorithm, it allows to find more possible solutions of the problem of node allocation. Using the developed simulator, we evaluate the utilization and average relative waiting job time of a job in a queue for systems up to 512 nodes with a 3D torus topology and up to 1024 nodes with a 4D torus topology. For all considered topologies the new node allocation algorithm, on average, improves the system utilization over the base algorithm by 9.61%. Similarly, for the average waiting job time relative to the requested job time, the gain is 2.04 times, on average. Secondly, we implement the ability to disable unused transceivers of the Angara interconnect router for each user job. Using the simulator and the same job workloads, the achieved energy saving is up to 2.78 kW for a system of 512 nodes (3D torus) and 4.41 kW for a system of 1024 nodes (4D torus).

Keywords: scheduling, torus topology, Angara, fine-grained job allocation, Slurm