14:20 - 14:40
Enhancing Data Center Efficiency and Throughput by Co-Scheduling
Paul Lin, Nestor Demeure, Mathias Wagner, Francesca Tartaglione, Garima Kochhar
NVIDIA, United States
Jonathan Skone, Dhruva Kulkarni, Jack Deslippe, Brian Austin
Lawrence Berkeley National Laboratory, CA, United States
Ian Karlin
Queen’s University, Canada
Many HPC data centers have supercomputers with separate CPU-only and GPU compute nodes to support different workloads. One approach to increase system throughput is to run CPU-only jobs on unused CPU resources on the GPU nodes. Analysis of the workload running on NERSC’s supercomputer showed that approximately 60% of Perlmutter’s GPU node-hours employed fewer than 12 CPU cores and less than 128GB host memory. The remaining 52 CPU cores and 128GB host memory could be used by co-scheduled CPU-only jobs. However, it is critical that the co-scheduled CPU-only jobs and GPU jobs have controlled/bounded interference. This preliminary study assesses the feasibility of co-scheduling in production. Testing of the NERSC-10 benchmark applications on NVIDIA Grace-Hopper nodes using Memory System Resource Partitioning and Monitoring (MPAM) and network/InfiniBand quality of service for performance isolation measured a 70% throughput increase from co-scheduling compared with no co-scheduling, while maintaining an acceptable level of performance isolation.
