The Embedded Genetic Allocator - A System to Automatically Optimize Mpi Communicator Mappings on High Performance Computing systems

摘要

This paper describes a new, automatic, protable approach to the problem of optimizing the mapping of processes within a parallelized software program to the individual processors in a high performance, scalable computing system. The resulting mapping optimizes the usage of the available inter-processor communications bandwidth and balances the computational processing load among the processors of the system, rivaling or even improving upon mappings generated by skilled programmers. The system consists of a genetic algorithm optimizer (the Embeddded Genertic Allocator or EGA), that reconfigures an inter-processor communication layer built upon the Message Passing Interface standard (MPI). The programmer writes the program using augmented MPI communicator calls and does not need to know the underlyign multiprocessor interconnection details. The EGA manipulates the allocations of processors to these communicators in an effort to minimize the execution time of time-critical portions of the target system. Trial allocations. Results are presented showing the EGA optimizing the performance of a conventional sonar signal processing chain hosted on a Mercury Computer RACEWAY.