Personalized MPI library for Exascale Applications and Environments

Minimizing the communication costs associated with a parallel application is a key challenge for the scalability of petascale and future exascale application. This paper introduces the notion of a personalized MPI library that is customized for a particular application and platform. The work is based on the Open MPI communication library, which has a large number of runtime parameters that can be modified without having to recompile the MPI library or the application. The approach described in this paper leads to a parameter set that minimizes the execution time of a given application. The paper presents a search algorithm adapted from experimental design theory to deal with large parameter spaces and the integration of the Open MPI mpiexec tool with a database to retrieve parameter sets based on application and platform specific settings. We demonstrate the benefits of this approach by tuning point-to-point and collective communication operations, and discuss advantages and disadvantages of various optimization strategies from an application developers point of view.