PIO-SIM

PIO-SIM is the layer of the simulator which simulates MPI-IO calls. In the case of collective I/O operations, the underlying implementation of the collective I/O algorithm is hidden from the user.

The basic idea behind collective I/O algorithms is to use global knowledge about the I/O transfer among the processes in order to combine many, small requests into fewer, large requests. PIO-SIM currently supports the following collective I/O implementations:

• Global Barrier Collective I/O: Upon making the collective I/O call, all processes execute a barrier synchronize at the outset, and then proceed with the regular, non-collective call. This form of collective I/O does not optimize the requests; it only serves as the base implementation that is allowed by the MPI-IO specification. This is the approach taken by versions 1.0 through 1.2 of MPI-IO/PIOFS, the implementation of MPI-IO by IBM for the SP2.

• Node Grouping: Processes participating in the collective call are partitioned into groups and the I/O requests for each group are implemented in turn. Thus node grouping attempts to avoid flooding the system with too many simultaneous requests. Its utility depends on the specific application, and in the best case, only an 8-times improvement in performance has been achieved [NF95].

• Two-Phase I/O: I/O is done in two phases [RBC93]. In the first phase, processes arrange their requests according to a conforming distribution. This is a permutation of the data across the processes so that it coincides as much as possible with the underlying file layout.

  The actual I/O is done in the second phase to reduce overall disk traffic. The I/O requests are optimized in two ways: 1) duplicate requests are eliminated, and 2) numerous small requests which lie contiguously are combined into one, large request.