Run-time support for asynchronous parallel computations

In this peeper we describe DREAM, a distributed environment that provides run-time support for parallel computations on asynchronous multiprocessors. The system supports global distributed arrays as collections of subarrays in the local memories of the intervening processors. Nodes allocate and deallocate array portions dynamically, and access external array sections without the intervention of the user programming in the remote node. Remote accesses can be performed while the program continues its execution, thus overlapping communication and computation. This feature allows the user to implement dynamic communication patterns by accessing external array elements on demand without incurring a heavy performance penalty. DREAM also vectorizes requests into larger network messages for efficiency. We report performance results for an application running on top of a prototype of the system, showing good scalability and masking the network latency with computation