Performance Optimization of PETSc on GPU-accelerated systems

PI Name Hong Zhang, MCS
PI Institution Argonne National Laboratory
Project Description

As the Aurora supercomputer will be built using GPUs, it is important for us to prepare PETSc for the upcoming challenges early. The goal of this project is to test PETSc’s configuration and build system on GPU, benchmark some selective linear algebra kernels in PETSc on GPU, identify the performance bottlenecks, and improve the workflow, data structures and algorithms to better utilize the GPU resources.

Testbed

KNL Nodes