SC16 Salt Lake City, UT

High Performance Geometric Multigrid (HPGMG): an HPC Performance Benchmark


Authors: Dr. Mark Adams (Lawrence Berkeley National Laboratory)

BP Abstract: This meeting facilitates community participation in the HPGMG project. HPGMG is a compact benchmark designed as a design tool for emerging architectures and a ranking metric. HPGMG is well balanced with respect to modern HPC applications and algorithms. We compile the HPGMG-FV list of the world’s largest supercomputers with the metric, a multigrid solve of a fourth order accurate finite volume discretization of the Laplacian. We released our first list at ISC16, and continue with our next biannual release at SC16. We encourage community participation with submissions to the HPGMG-FV list, and contributed talks and discussion in the BOF.

Long Description: This is the third SC BoF for the HPGMG project. We presented the first official HPGMG-FV list at ISC this past June and present the next list in this BoF. In this past year we have stabilized the metric specification on a fourth order accurate discretization of the 3D Laplacian on a Cartesian grid. We provide analysis to gain insights regarding the architecture and performance characteristics of top500 machines. See https://hpgmg.org for current status and news. This year we aim for stimulating a discussion of benchmarking issues with speakers from the HPGMG team, metric users and submitters to the HPGMG database, and researchers on topics relevant to HPGMG, followed by an open discussion with the audience and speakers. Users of HPGMG, and researchers in benchmarking and advance programming models interested in speaking are encouraged to contact us. Centers interested in submitting results are encouraged to visit our web page (http://crd.lbl.gov/projects/hpgmg) and the repository (https://bitbucket.org/hpgmg/hpgmg). This project is motivated by the loss of effectiveness of the HPL benchmark as a proxy for a wide variety of application relevant to the HPC community, although HPL continues to be an effective proxy for applications based on dense linear algebra. HPL benchmark is the most widely recognized metric for ranking high-performance computing systems. When HPL gained prominence in the early 1990s there was a strong correlation between its predicted ranking of a system and the efficacy of the system for full-scale applications. Computer system vendors pursued designs that would increase HPL performance, which would in turn improve overall application performance. This has ceased to be the case and in fact the opposite is now true. HPL rankings of computer systems are no longer well correlated to real application performance, which use more work optimal algorithms with high bandwidth and low latency requirements. HPGMG is designed to have machine sensitivities that correlate well with the sensitivities of HPC applications. We will begin with short talks from the HPGMG developers, who discuss the HPGMG-FV metric, and provide performance data and analysis. We will continue with talks from users of HPGMG and researchers in the benchmarking community, including talks on Nvidia’s Pascal and Intel’s Knights Landing optimizations. This is our first BoF with KNL performance data and tuning results. Interested speakers are invited to contact us about giving a short presentation. We conclude with an open forum and discussion with the audience and speakers.

Conference Presentation: pdf


Birds of a Feather Index