Implementation and Performance of the MPI

MPI is the new standard which deenes a set of message passing operations for multicomputers and clustered systems. In comparison to other popular message passing systems, MPI provides a richer collection of functions, allowing eecient implementations , portability and excellent support for the development of parallel libraries. In this paper, we describe the implementation and performance of MPI on the Fujitsu AP1000 multicomputer. To produce an eecient implementation, the operating system on the AP1000 had to be modiied to better support MPI. These modiications are presented , along with the hardware operations that were utilised. A selective broadcast operation was developed from the modiications which allowed very ee-cient group-wide broadcast. The performance of the implementation in comparison to native AP1000 calls is presented with benchmarks of the collective routines implemented using the selective broadcast operation. 1 The MPI Standard The message passing paradigm has been used for many years within distributed memory parallel machines and clustered workstation environments. Due to the paradigm's popularity, a plethora of message passing systems currently exist, and until recently, there was no standard message passing interface. The situation changed in April 1994, when the MPI Forum released the nal version of the MPI standard 3]. The inception of the standard allows the development of portable message passing code and deenes many advanced features to assist in building parallel libraries. One of the goals of the standard is to allow eecient implementations, and as a result the standard is quite large (129 functions deened) and complex. Communication in MPI is performed within user-constructed process groups, which specify a subset of processes in a parallel machine. Communication within a group occurs in a separate communication universe or context. Furthermore, every communication belongs in some speciied context. Contexts are a set of system-managed tags which partitions the communications space, completely separating messages that are sent in one context from operations (such as receive-message) that specify another context. Process groups and contexts together form what is known in MPI as an intracommunicator, which consists of a set of processes and a unique context that the group members communicate with. Processes are uniquely identiied within an intracom-municator by a rank, which ranges from 0 : : : n ?1, where n is the size of the group. The insulating property of contexts is important when building safe and robust parallel libraries, which was one of MPI's major design objectives 8]. …