HPCG Benchmark Update
Authors: Dr. Michael Heroux (Sandia National Laboratories)
Abstract: The High Performance Conjugate Gradients (HPCG) Benchmark is a community metric for ranking HPC systems. The first list of results was released at ISC'14, including optimized results for systems built upon Fujitsu, Intel, Nvidia technologies. Lists have been announced at SC14, IS'14, SC15 and ISC16, with an increase from 15 to 25, 40, 64 and 82 entries, respectively.
In this BOF we present an update of HPCG 3.1, and opportunities for optimizing performance, with presentations from all vendors who have participated in HPCG optimization efforts. We spend the remaining time in open discussion.
Long Description: The High Performance Conjugate Gradients (HPCG) Benchmark is designed to test features of a high performance computing (HPC) system in a way that complements the high performance Linpack (HPL) benchmark. HPL tends to approach the maximum achievable floating point performance on a given system, while most real applications reach a tiny fraction of what HPL achieves. In contrast, HPCG tests interconnect latency and bandwidth, and on-node concurrency capability of multicore, manycore, accelerator and heterogeneous processors.
HPL derives its performance from an existing collection of optimized dense matrix kernels. As a result, achieving good performance from HPL is fairly straightforward. HPCG is a newer benchmark that depends primarily on sparse linear algebra and optimization strategies and implementations are still emerging. Even so, Fujitsu, IBM, Intel, Nvidia and their related integrators and leadership computing facilities have made significant investments in HPCG optimization, obtaining substantial performance improvements.
In this BOF we review the architecture of the HPCG reference code, emphasizing opportunities for improving its performance and describing what kinds of optimizations are permissible, especially for the latest version HPCG 3.1. We follow this with a presentation from each of the computer system vendors on how they have optimized HPCG for their systems. We conclude the BOF with a general discussion about HPCG, future plans for its design and implementation and questions. At the very end we intend to announce the latest rankings and give awards for the top 3 systems.
This BOF will be valuable to any HPC community member who is interested in benchmarking and how HPCG can be used to obtain understanding about large-scale system performance.
Birds of a Feather Index