Abstract

Achieving maximum performance in message-passing programs requires that calculation and communication be overlapped. However, the program transformations required to achieve this overlap are error-prone and add significant complexity to the application program. We argue that calculation/communication overlap can be achieved easily and consistently by executing multiple threads of control on each processor, and that this approach is practical on message-passing architectures without any special hardware support. We present timing data for a typical message-passing application, to demonstrate the advantages of our scheme. 1 Introduction The most natural way to program many scientific algorithms on message-passing machines is in the loosely synchronous, or coarse-grain data parallel, style. Such programs alternate between phases of pure calculation and pure communication. This approach is easy for the programmer to understand, so the resulting programs are relatively simple to debug and ...

Citations