Workshops Archive

HPL and HPL-MxP at Exascale on Aurora: Performance Engineering and Characterization


Workshop: PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems

Authors: Kazushige Goto and Huda Ibeid (Intel Corporation), Kalyan Kumaran and Servesh Muralidharan (Argonne National Laboratory), and Anthony-Trung Nguyen and Aditya Nishtala (Intel Corporation)

Abstract: Running HPL and HPL-MxP at exascale requires more than optimized numerical kernels; it requires coordinated performance engineering across communication, resource mapping, runtime configuration, CPU--GPU interaction, and mixed-precision execution. This paper reports experience from large-scale benchmarking campaigns on Aurora, the Intel-based exascale supercomputer at the Argonne Leadership Computing Facility. Aurora progressed from 0.585 EF/s on 5,439 nodes to 1.01 EF/s on 9,234 nodes in FP64 HPL, while HPL-MxP reached 11.64 EF/s on 9,500 nodes. We describe the system-level choices behind these results, including locality-aware rank, GPU, memory, and NIC placement; explicit CPU--GPU pipelining; host-buffer MPI and transport configuration; phase-specific NIC assignment; AMX-accelerated mixed precision; and a hybrid point-to-point/collective path introduced after synchronization stalls at scale. This experience-based study concludes with practitioner guidance separating Aurora-specific findings from broader principles for tightly coupled heterogeneous systems.


Back to PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems Archive Listing

Back to Full Workshop Archive Listing