Preliminary I/O Performance Evaluation on GPU Accelerator and External Memory

Bibliographic Information

Other Title
  • GPUアクセラレータと不揮発性メモリを考慮したI/O性能の予備評価

Search this article

Abstract

Recent supercomputers deploy not only many-core accelerators such as GPU, Xeon Phi, but also Non-Volatile RAM such as flash memory as an external memory, in order to handle large-scale data processing for a wide range of applications. In particular, aggregating multiple mini SATA SSDs (mSATA SSDs) has potential to perform high bandwidth, high IOPS as well as large capacity. However, how to exploit I/O performance from multiple mSATA SSDs is not clear, especially when combining with many-core accelerators. In order to clarify I/O characteristics between GPU and multiple mSATA SSDs, we comparatively investigate I/O strategies on multiple mSATA SSDs and GPU. Our preliminary results exhibit that sequential read bandwidth is 7.69GB/s and write bandwidth is 3.8GB/s. Our results also exhibit 3.06GB/s of throughput from 8 mSATA SSDs to GPU by using RAID0 and setting stripe size to 1MB. Using pinned memory and data chunk size also affect data transfer performance significantly.

Journal

  • IPSJ SIG Notes

    IPSJ SIG Notes 2013 (1), 1-9, 2013-09-23

    Information Processing Society of Japan (IPSJ)

Keywords

Details

  • CRID
    1571135652920954112
  • NII Article ID
    110009606422
  • NII Book ID
    AN10463942
  • Text Lang
    ja
  • Data Source
    • CiNii Articles

Report a problem

Back to top