Abstract
The NERSC Exascale Science Application Program (NESAP) program has two early users from the ACME project: the MPAS Ocean model (Jacobsen) and the full ACME model (Johansen, Keen, Taylor). As early users, ACME will get early access and porting help, specifically in preparation for the Cori "Phase 2" machine, which will be coming online in Oct 2016. Cori Phase 1 has 3,260 Intel Haswell processors; Phase 2 will have 9,300 Intel "KNL" processors, which are expected to have 3+ TeraFLOPS double-precision peak performance, for more than 30 PetaFLOPS of computing capacity. However, the architectural changes in Cori have implications for ACME code: AVX512 vector registers must be used effectively by code to achieve peak FLOPS, and a more complex memory hierarchy (including use of its unique "high-bandwidth" MCDRAM memory) will require some code modifications. In addition, Cori will have the ability to use Burst Buffers to improve I/O performance and perform some "in situ" analysis, if ACME components can take advantage of it. In this poster, we will provide an architecture overview of the new Cori platform, and specific examples of ACME code changes that improved performance. In addition, we will describe future opportunities and risks in running ACME on NERSC's Cori platform going forward.
NOTE: Please contact Hans Johansen (Unlicensed) (hjohansen@lbl.gov) with questions.