This benchmark measures the speed at which large amounts of data can be written to disk using the IOHDF5 I/O method and the HDF5 file format. It does that by generating an initial data checkpoint of the simulation, saving the current contents of 100 3-dimensional grid functions of the application. These grid functions are all given the local size on each processor of 80-cubed, and use REAL*8 data types, so that each processor contributes a total of 352 MBytes to the overall initial data checkpoint.
At the moment this benchmark contains parameter files for different I/O modes, each one with a different level of parallelism when doing output: with onefile the first processor receives data from all other processors and then does all the output to disk; 8proc uses one I/O processor per group of 8 processors; eachproc writes one (chunked) checkpoint file per processor.
These files are also contained in the benchmark’s tarball.
For an overview of the commands on how to download, compile and run the benchmark see Downloading and building the benchmark tarballs.