An Empirical Implementation of I/O Separation Scheme for Burst Buffers in HPC Systems

Abstract

To meet the exascale I/O requirements in the High-Performance Computing (HPC), a new I/O subsystem, named Burst Buffer, based on non-volatile memory, has been developed. However, the diverse HPC workloads and the bursty I/O pattern cause severe data fragmentation to SSDs, which creates the need for expensive garbage collection (GC) and also increase the number of bytes actually written to SSD. The new multi-stream feature in SSDs offers an option to reduce the cost of garbage collection. In this paper, we leverage this multi-stream feature to group the I/O streams based on the user IDs and implement this strategy in a burst buffer we call BIOS, short for Burst Buffer with an I/O Separation scheme. Furthermore, to optimize the I/O separation scheme in burst buffer environments, we propose a stream-aware scheduling policy based on burst buffer pools in workload manager and implement the real burst buffer system, BIOS framework, by integrating the BIOS with workload manager. We evaluate the BIOS and framework with a burst buffer I/O traces from Cori Supercomputer including a diverse set of applications. We also disclose and analyze the benefits and limitations of using I/O separation scheme in HPC systems. Experimental results show that the BIOS could improve the performance by 1.44× on average and reduce the Write Amplification Factor (WAF) by up to 1.20×, and prove that the framework can keep on the benefits of the I/O separation scheme in the HPC environment.