The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Workshop:PDSW: Sixth International Parallel Data Systems Workshop
Authors: Glenn K. Lockwood, Alberto Chiusole, and Nicholas J. Wright (Lawrence Berkeley National Laboratory (LBNL))
Abstract: The proliferation of extreme-scale analytics and AI have motivated the creation of new parallel storage systems that either use all flash or combine nonvolatile memory (NVRAM) and flash to achieve the best attributes of both media. Clearly such storage systems have relevance in HPC, especially to meet the increasing needs of data analytics and AI in scientific computing, and evaluating such technologies for their applicability to HPC workloads is an important goal. In this paper we present a preliminary evaluation of the VAST storage system which incorporates NVRAM and QLC flash along with advanced technologies and algorithms to manage data between these two media. Our results show that traditional HPC performance measurement techniques struggle to properly characterize such architectures because they were designed to test much simpler storage systems. We discuss these shortcomings and propose alternative methods that better reflect the performance that real workflows can expect.
Back to PDSW: Sixth International Parallel Data Systems Workshop Archive Listing