Performance Comparison of GPFS 1.3 and GPFS 1.4 for POSIX and MPI-IO

Hedges, R ; Loewe, W

Lawrence Livermore National Laboratory

关键词: 99 General And Miscellaneous//Mathematics, Computing, And Information Science;    Parallel Processing;    G Codes;    Comparative Evaluations;    Performance;   

DOI  :  10.2172/15007273
RP-ID  :  UCRL-ID-144548
RP-ID  :  W-7405-ENG-48
RP-ID  :  15007273

美国|英语

来源: UNT Digital Library

【 摘 要 】

This report by the SIOP observes the effects of recent hardware and software changes for parallel I/O performance to the GPFS parallel file system. The IBM SP machine (frost) has been upgraded from Mohonk with GPFS 1.3 to Mohonk2 with GPFS 1.4. In addition, the Colony switch adapters have been upgraded from Single/Single to Double/Single. The tests discussed here were performed on frost using 60 compute nodes and the GPFS file system using 2 dedicated I/O nodes (servers). The tests performed utilize the POSIX and MPI-IO interfaces to GPFS. The noted system changes to frost have improved both POSIX and MPI-IO peak read performance and have not diminished peak write performance. We note that as the bandwidth of mounted disks is near fully utilized, there was no expectation of significant performance improvement. For POSIX, the best write rates did not change from 550 MB/sec. The read rates improved from 500 MB/sec to 600 MB/sec, however. For MPI-IO, the best write rates did not change from 550 MB/sec. The read rates improved from 470 MB/sec to 570 MB/sec, in line with the improvements observed. The MPI-IO discontiguous test results show that improvement is significant (nearly a factor of 2 beyond 40 nodes). The performance of this particular test is more sensitive to the improved switch performance characteristics because the data passes across the switch twice: once in the MPI-IO datashipping phase to assemble large block; and then to write the data out to the disks as this requires two passes across the switch.