This paper will be presented at SC19 in Denver, Colorado, USA in the Improving Next-Generation Performance and Resilience session on Thursday 21 November 2019, 2.00-2.30 p.m., Location 405-406-407.
It will subsequently be published in the Proceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis.
Abstract: Memory and I/O performance bottlenecks in supercomputing sim- ulations are two key challenges that need to be addressed on the road to Exascale. The recently released byte-addressable persistent non-volatile memory technology from Intel, DCPMM, promises to be an exciting opportunity to break with the status quo, with unprecedented levels of capacity at near-DRAM speeds. In this paper, we explore the potential of DCPMM in the context of high- performance scientific computing using two distinct applications in terms of outright performance, efficiency and usability for both its Memory and App Direct modes. In Memory mode, we show that it is possible to achieve equivalent performance and better efficiency for a CASTEP simulation that struggles with memory capacity limitations on conventional DRAM-only systems without needing to introduce any changes to the application. For IFS, we demonstrate that using a distributed object-store over the NVRAM devices reduces the data contention created in weather forecasting data producer-consumer workflows. In addition to presenting the impact on two applications, we also present results for achievable memory bandwidth performance using STREAM.