Abstract | ||
---|---|---|
I/O performance remains a weakness of parallel computingsystems today. While this weakness is partly attributedto rapid advances in other system components, I/Ointerfaces available to programmers and the I/O methodssupported by file systems have traditionally not matchedefficiently with the types of I/O operations that scientific applicationsperform, particularly noncontiguous accesses.The MPI-IO interface allows for rich descriptions of theI/O patterns desired for scientific applications and implementationssuch as ROMIO have taken advantage of thisability while remaining limited by underlying file systemmethods.A method of noncontiguous data access, list I/O, wasrecently implemented in the Parallel Virtual File System(PVFS). We implement support for this interface in theROMIO MPI-IO implementation. Through a suite of non-contiguousI/O tests we compared ROMIO list I/O to currentmethods of ROMIO noncontiguous access and foundthat the list I/O interface provides performance benefits inmany noncontiguous cases. |
Year | DOI | Venue |
---|---|---|
2003 | 10.1109/CCGRID.2003.1199358 | CCGrid |
Keywords | Field | DocType |
o performance,o accesses,o operation,mpi-io interface,noncontiguous case,romio noncontiguous access,o pattern,romio list,noncontiguous access,o interface,noncontiguous data access,parallel processing,mathematics,testing,computer science,distributed computing,message passing,data access | Virtual file system,File system,Suite,Computer science,Parallel processing,Parallel computing,Input/output,Programming profession,Data access,Operating system,Message passing,Distributed computing | Conference |
ISBN | Citations | PageRank |
0-7695-1919-9 | 48 | 2.72 |
References | Authors | |
12 | 6 |
Name | Order | Citations | PageRank |
---|---|---|---|
Avery Ching | 1 | 221 | 16.21 |
Alok N. Choudhary | 2 | 3441 | 326.32 |
Kenin Coloma | 3 | 137 | 9.36 |
Wei-keng Liao | 4 | 1095 | 87.98 |
Robert Ross | 5 | 2717 | 173.13 |
William D. Gropp | 6 | 5547 | 548.31 |