Workshop: HPC Systems Professionals Workshop (HPCSYSPROS26)
Authors: Daniel Margala, Jonathan Skone, and Christopher Samuel (LBNL NERSC)
Abstract: Heterogeneous systems such as NVIDIA GH200 and GB200 pair powerful GPUs with substantial CPU capacity, but GPU jobs targeting these systems frequently leave much of the CPU capacity idle, lowering CPU utilization and overall system throughput. Co-scheduling CPU workloads alongside GPU jobs improves resource utilization, but can introduce interference and undermine performance predictability. We present a proof of concept for co-scheduling CPU and GPU jobs on GH200 nodes with hardware-enforced resource isolation using Arm Memory System Resource Partitioning and Monitoring (MPAM) for CPU, L3-cache, and memory bandwidth and InfiniBand QoS for host network traffic. We augment existing resource isolation mechanisms in Slurm to allocate jobs to dedicated CPU and GPU "shards" with allocations for compute cores, L3-cache, memory bandwidth, and network bandwidth. This proof-of-concept implementation provides a practical framework for evaluating co-scheduled workloads with hardware-enforced resource isolation and establishes an operational model for multi-tenant co-scheduling on heterogeneous HPC/AI nodes.
Back to HPC Systems Professionals Workshop (HPCSYSPROS26) Archive Listing
Back to Full Workshop Archive Listing