Tuning the I/O performance of PHDF5 is both application and system dependent. There are several parameters that one can tune, ranging from parallel file system parameters to MPI-IO and HDF5 parameters. As this topic is about PHDF5, only HDF5 tuning parameters are covered.

  • Metadata Cache: You can adjust the size of metadata cache to avoid small writes using the H5Pset_mdc_config() function from the API.
  • Chunking: Using chunked dataset storage, application I/O performance can vary based on the chunk size and the block boundaries of the file system. Users may need to experiment a bit to find the chunk size that leads to good I/O performance.
  • Others: Trading off between collective and independent I/O can also be treated as a design choice to improve I/O.
 
©  |   Cornell University    |   Center for Advanced Computing    |   Copyright Statement    |   Inclusivity Statement