| | 1 | == Benchmark programs == |
| | 2 | * IOR (http://sourceforge.net/projects/ior-sio) has a 'driver' for pnetcdf |
| | 3 | * The FLASH I/O kernel (http://www-unix.mcs.anl.gov/pio-benchmark/code/flash-argonne-20060913.tar.gz) will measure several write-only workloads |
| | 4 | |
| | 5 | |
| | 6 | == Notes and suggestions == |
| | 7 | |
| | 8 | Be mindful of benchmarks which write large contiguous chunks of data. While |
| | 9 | interesting, it might not be what a scientific application using pnetcdf would |
| | 10 | actually do. |
| | 11 | |
| | 12 | Older versions of IOR triggered a performance bug on some Lustre file systems. |
| | 13 | Sometimes slow write performance on Lustre can be fixed by disabling the "data |
| | 14 | sieving" optimization. See HintsForPnetcdf for more information. |