MPICH-G: An MPI for Grids

CLUSTERWORLD volume 2 no 11 3 volume 2 no 11 CLUSTERWORLD In January , researchers from more than  organizations from research laboratories, academia, and industry formed the Message Passing Interface Forum. e intention was to define a set of library interfaces, with the goal of producing a widely used standard for message-passing programs. In June , the forum published version . of the Message Passing Interface (MPI) standard. As the standard was being developed, members of the forum at Argonne National Laboratory and Mississippi State University began implementing MPI in a library they called MPICH. Early on, they decided to internally split their library into two parts: a “top” part including everything found in the standard (e.g., all the MPI functions application programs call) and a “bottom” part comprising a much smaller set of functions primarily responsible for transporting data. Functions in the top part perform all the parameter checking and implement some MPI functions in terms of others (e.g., MPI_Bcast implemented in terms of MPI_Send and MPI_Recv). Functions in the bottom part focus exclusively on point-to-point communication. To define the interaction between the top and bottom parts of the library, the MPICH developers created the Abstract Device Interface (ADI). e two-layer design and the new ADI provided a framework in which the standard could be implemented quickly. Industrial developers (HP, SGI, and Intel, to name just a few) were able to produce their own MPI libraries by first acquiring the top portion of the MPICH library, freely distributed by the MPICH team soon after the standard was published, and then concentrating their efforts MPICH-G: An MPI for Grids