• H. Shang, M. Borland, L. Emery, R. Soliday, Y. Wang
  ANL, Argonne

Use of SDDS, the Self-Describing Data Sets file protocol and toolkit, has been a great benefit to development of several accelerator simulation codes. However, the serial nature of SDDS was found to be a bottleneck for SDDS-compliant simulation programs such as parallel elegant. A parallel version of SDDS would be expected to yield significant dividends for runs involving large numbers of simulation particles. In this paper, we present a parallel interface for reading and writing SDDS files. This interface is derived from serial SDDS with minimal changes, but defines semantics for parallel access and is tailored for high performance. The underlying parallel IO is built on MPI-IO. The performance of parallel SDDS and parallel HDF5 are studied and compared. Our tests indicate better scalability of parallel SDDS compared to HDF5. We see significant I/O performance improvement with this parallel SDDS interface.