GPU Behavior on a Large HPC Cluster

摘要

We discuss observed characteristics of GPUs deployed as accelerators in an HPC cluster at Los Alamos National Laboratory. GPUs have a very good theoretical FLOPS rate, and are reasonably inexpensive and available, but they are relatively new to HPC, which demands both consistently high performance across nodes and consistently low error rate. We modified a standard acceptance procedure to test GPU performance, error rate and reliability characteristics, and ran the test suite on a Fermi HPC cluster at LANL. We discuss here our methodology for this testing, and present results relevant to the deployment of GPUs in an HPC environment. In this paper we show performance variability, power usage variability (possibly related), and some reliability concerns on the GPUs tested. We argue for rigorous testing of these devices in deployment as a way of characterizing their behavior.