Parallel IO Support for Meta-computing Applications: MPI_Connect IO Applied to PACX-MPI

Parallel IO (PIO) support for larger scale computing is becoming more important as application developers better understand its importance in reducing overall execution time by avoiding IO overheads. This situation has been made more critical as processor speed and overall system size has increased at a far greater rate than sequential IO performance. Systems such as MPI_Connect and PACX-MPI allow multiple MPPs to be interconnected, complicating IO issues further. MPI_Connect implemented Parallel IO support for distributed applications in the MPI_Conn_IO package by transferring complete sections of files to remote machines, supporting the case that all the applications and the file storage were completely distributed. This system had a number of performance drawbacks compared to the more common usage of metacomputing where some files and applications have an affinity to a home site and thus less data transfer is required. Here we present the new PACXMPI PIO system based initially on MPI_Connect IO, and attempt to demonstrate multiple methods of handling MPI PIO that cover a greater number of possible usage scenarios. Given are some preliminary performance results as well as a comparison to other PIO grid systems such as the Asian Pacific GridFarm, and Globus gridFTP, GASS and RIO.