Goal is to profile several codes of increasing complexity. Some material on using Darshan on Archer2 should be produced.
3D array grid and profile with darshan trough different optimizations
- POSIX write per process
- MPI-IO independent ( suffers from file locking ? )
- MPI-IO collective ( file locking issue eased )
- Changing the number of stripes ( OST bandwith limit)
- MPI-IO HDF5 ( similar scaling as MPI-IO but with an overhead from many small metadata writes )
- MPI-IO Adios2 ( scaling should be much improved )
Possible other factors to explore
- tuning environment variables
- looking at verbose output of MPI-IO
- write 1D contigous array ( no file locking issues )
- use MPICH_MPIIO_STATS, MPICH_MPIIO_TIMERS, MPICH_MPIIO_HINTS_DISPLAY
- Documentation on how to use darshan
- Slides for following throughout the stack