Exploiting the Performance Benefits of Storage Class Memory for HPC and HPDA Workflows

Abstract

Byte-addressable storage class memory (SCM) is an upcoming technology that will transform the memory and storage hierarchy of HPC systems by dramatically reducing the latency gap between DRAM and persistent storage. In this paper, we discuss general SCM characteristics, including the different hardware configurations and data access mechanisms SCM is likely to provide. We outline the performance challenges I/O requirements place on traditional scientific workflows and present how data access through SCM can have a beneficial impact on the performance of such workflows, in particular those with large scale data dependencies. We describe the system software components that are required to enabled workflow and data aware resource allocation scheduling in order to optimise both system throughput and time to solution for individual applications; these include a data scheduler and data movers. We also present an illustration of the performance improvement potential of the technology, based on initial workflow performance benchmarks with I/O dependencies.