A team of researchers funded by the Exascale Computing Project demonstrated the efficacy of combining DRAM and high-density, byte-addressable nonvolatile memory (NVM) for accelerating high-performance computing (HPC) and enabling large problems at reasonable performance beyond the capabilities of DRAM and NVM alone. They also developed a prediction model to investigate how HPC applications could exploit NVM for increased capacity while circumventing the limitations of latency, bandwidth, and durability. Their findings were published in the Proceedings of the 2020 IEEE International Parallel and Distributed Processing Symposium.
High-density NVM has emerged as a memory/storage tier for exascale computing. It enables larger memory capacity than DRAM, and recent developments may provide bandwidth and latency comparable to DRAM; however, because NVM is still under active development, the two are typically paired to provide optimized memory performance.
The researchers used the Seven Dwarfs, well-known sequential and parallel algorithmic methods, and flagship libraries such as ScaLAPACK, SuperLU, and Hypre with the first large-scale, commercially available NVM hardware to evaluate the efficacy of DRAM-NVM heterogeneous memory with dense and sparse linear algebra, spectral methods, N-body methods, structured and unstructured grids, and Monte Carlo–based algorithms. They experimented with data placement and new memory options and configurations and identified write throttling and concurrency control as areas for optimizing HPC applications with NVM-based main memory. The team discovered, for example, that increased concurrency provides increased read bandwidth but decreased write bandwidth; their work suggested that existing applications can be optimized through concurrency control and write-aware data placement.
Peng, Ivy, Kai Wu, Jie Ren, Dong Li, and Maya Gokhale. 2020. “Demystifying the Performance of HPC Scientific Applications on NVM-Based Memory Systems.” 2020 IEEE International Parallel and Distributed Processing Symposium (IPDPS) (May). doi:10.1109/ipdps47924.2020.00098. http://dx.doi.org/10.1109/IPDPS47924.2020.00098.
