Management of Communication Environments for Minimally Synchronous Parallel ML

Minimally Synchronous Parallel ML is a functional parallel language whose execution time can then be estimated and dead-locks and indeterminism are avoided. Programs are written as usual ML programs but using a small set of additional functions. Provided functions are used to access the parameters of the parallel machine and to create and operate on a parallel data structure. It follows the cost model of the Message Passing Machine model (MPM). In the current implementation, the asynchrony is limited by a parameter called the asynchrony depth. When processes reach this depth a global synchronization occurs. This is necessary to avoid memory leak. In this paper we propose another mechanism to avoid such synchronization barriers. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.