Workshop: Sustainable Supercomputing
Authors: Zhong Zheng (University of Illinois at Chicago); Michael E. Papka (Argonne National Laboratory, University of Illinois at Chicago); and Zhiling Lan (University of Illinois at Chicago, Argonne National Laboratory)
Abstract: Power constrained HPC systems increasingly execute heterogeneous CPU–GPU applications under strict cluster wide power limits. Existing power management policies typically rely on fair sharing or utilization based heuristics, overlooking application specific sensitivity to CPU and GPU power caps and therefore allocating reclaimed power inefficiently. We present EcoShift, a performance aware framework for cluster wide power management. EcoShift combines online performance prediction with a dynamic programming allocator that distributes reclaimed power among CPU–GPU applications to maximize average performance improvement. We evaluate EcoShift through emulation on two heterogeneous platforms equipped with Intel CPUs and NVIDIA A100 or H100 GPUs across diverse CPU–GPU workloads. EcoShift consistently outperforms state of the art policies, delivering up to 6% higher average performance while strictly maintaining the cluster wide power constraint.
Back to Sustainable Supercomputing Archive Listing
Back to Full Workshop Archive Listing