Performance testing and tuning of CkIO
- getting preferred block sizes from FS-specific (e.g. Lustre, GPFS) metadata
- node mapping for proximity to IO nodes to get data out of the interconnect
O_DIRECTon Lustre to avoid locking
- performance of many small writes (potential for optimization using TRAM)
#5 Updated by Eric Bohm over 5 years ago
Comparison (and/or usage of) the GLEAN framework should also be considered (http://www.mcs.anl.gov/project/glean-situ-visualization-and-analysis). It has been used to get close to peak performance out of pnetcdf and HDF5 on BlueGene architectures, which is pretty impressive given how unremarkable HDF5 and PnetCDF are otherwise.