Management of Communication Environments for Minimally Synchronous Parallel ML

Abstract

Minimally Synchronous Parallel ML is a functionalparallel language whose execution time can then beestimated and dead-locks and indeterminism areavoided. Programs are written as usual ML programsbut using a small set of additionalfunctions. Provided functions are used to access theparameters of the parallel machine and to create andoperate on a parallel data structure. It follows thecost model of the Message Passing Machine model(MPM). In the current implementation, the asynchronyis limited by a parameter called the asynchronydepth. When processes reach this depth a globalsynchronization occurs. This is necessary to avoidmemory leak. In this paper we propose a mechanism toavoid such synchronization barriers. This mechanismrelies on a more complex management of thecommunication environments but with a small andparametrized overhead.