The National Energy Research Scientific Computing Center (NERSC) at Lawrence Berkeley National Laboratory has been at the leading edge of HPC architecture and deployment for 45 years. Advances in extreme-scale storage coming from the enterprise and hyperscale industries are finding their way into HPC systems to address challenges stemming from massive datasets and AI-based workloads, but fundamental differences in I/O patterns are leaving gaps. In this talk, we will compare the storage and I/O requirements of large-scale HPC workloads with those of the cloud and show how HPC’s unique requirements have led NERSC to deploy NVMe in the form of burst buffers and all-flash parallel file systems rather than block- and object-based storage. We will then explore how recent technological advances that target enterprise and cloud I/O workloads may also benefit HPC, and we will highlight a few remaining challenge areas in which innovation is required.

About the Speaker

Glenn K. Lockwood is a storage architect for NERSC at Lawrence Berkeley National Laboratory who specializes in holistic I/O performance analysis, extreme-scale storage architectures, and emerging I/O technologies. He heads the I/O performance efforts for the 30 PB all-flash parallel file system being deployed as part of the upcoming Perlmutter system in 2020 and leads the pathfinding efforts in storage and I/O for NERSC’s 2024 exascale system. Glenn is also a maintainer of the industry-standard IOR and mdtest I/O benchmarks and contributes to the Darshan I/O profiling library. Glenn previously worked in the genomics industry and has a formal research background in silica surface chemistry. He holds a Ph.D. in materials science and B.S. in ceramic engineering, both from Rutgers University.