PRFX: a runtime library for high performance programming on clusters of SMP nodes

Publisher Summary This chapter discusses PRFX, a library for implementing parallel algorithms with both static and dynamic properties and executing them on clusters of symmetric multiprocessor (SMP) nodes. The parallelism is expressed through RPC (tasks) working on special shared data. An inspector builds the task DAG, which is used to schedule and guide the parallel execution. The implementation of PRFX uses POSIX threads with one-sided communications. The chapter discusses PRFX by experimental results for LU matrix factorization and Laplace equation solver on an IBM SP3. The chapter describes a runtime library for clusters of SMP, dedicated to a specific class of parallel algorithms having static and dynamic properties. The inspection phase builds the task DAG for a specific parallel execution by using a set of structuring data. Then a scheduling algorithm is applied to this task DAG in order to get the task vector of each executor thread. This step is done by a single CPU, while all others are waiting the end of the inspection. The RPC tree must be built and fixed from a partial execution (inspection) of the code on an input set of structuring data.