Participant : Luiz Angelo Steffenel.
LaPIe is an automatically tuned collective communication library designed for large-scale heterogeneous systems.
The popularity of heterogeneous parallel processing environments like clusters and computer grids has emphasised the impact of network heterogeneity on the performance of parallel applications. Collective communication operations are especially concerned by this problem, as heterogeneity interfers directly on the communication performance.
LaPIe provides a set of MPI collective communication operations especially designed to perform automatic adaptation according to the network characteristics. Indeed, LaPIe combines both topology discovery and performance prediction to chose the best communication scheduling that minimizes the overall communication time for a given operation.
LaPIe is distributed as a programming library that overloads (through profiling) existing MPI calls. This allows LaPIe to be easily integrated into existing applications without modifying their code - we just need to recompile them. In addition, LaPIe was designed to facilitate the addition of new scheduling techniques and collective communication operations. Therefore, LaPIe provides an excelent testbed to develop and evaluate new communication scheduling techniques and/or architecture-specific optimizations.
LaPIe currently supports MPI_Bcast , MPI_Scatter and MPI_Alltoall operations, whose efficiency was evaluated in several papers we published. New operations such as MPI_Reduce , MPI_Gather and MPI_Allgather should be added very soon.
LaPIe is available at http://libresource.inria.fr/projects/LaPIe .