ASYNC Loop Constructs for Relaxed Synchronization

Abstract

Conventional iterative solvers for partial differential equations impose strict data dependencies between each solution point and its neighbors. When implemented in OpenMP, they repeatedly execute barrier synchronization in each iterative step to ensure that data dependencies are strictly satisfied. We propose new parallel annotations to support an asynchronous computation model for iterative solvers. At the outermost level, the ASYNC_REDUCTION keyword is used to annotate the iterative loop as a candidate for asynchronous execution. The ASYNC_REGION contains inner loops which may be annotated by ASYNC_DO or ASYNC_REDUCTION. If the compiler accepts the ASYNC_REGION designation, it converts the iterative loop into a parallel section executed by multiple threads which divide the iterations of each ASYNC_DO or ASYNC_REDUCTION loop and execute them without having to synchronize through a conventional barrier. We present experimental results to show the benefit of using ASYNC loop constructs in multigrid methods and an SOR-preconditioned CG solver.