An empirical study of I/O separation for burst buffers in HPC systems

To meet the exascale I/O requirements for High-Performance Computing (HPC), a new subsystem, Burst Buffer, based on solid state drives (SSD), has been developed. However, diverse HPC workloads and bursty pattern cause severe data fragmentation that requires costly garbage collection (GC) increases number of bytes written to SSD. address this challenge, multi-stream feature developed SSDs. In work, we develop an Separation scheme called BIOS leverage group streams user IDs. We propose stream-aware scheduling policy burst buffer pools in workload manager, integrate with manager optimize separation buffer. evaluate proposed framework traces from Cori Supercomputer including set applications. Experimental results show could improve performance by 1.44x average reduce Write Amplification Factor (WAF) up 1.20x. These demonstrate potential benefits storage systems.