Execution of Regular DO Loops on Asynchronous Multiprocessors

This paper studies issues concerning parallel execution of regular Fortran DO loops on an asynchronous shared-memory multiprocessor, where each iteration is the basic unit to be executed by a single processing element. An iteration is a dependent predecessor of another iteration if execution of the latter iteration has to wait until execution of the former iteration has completed. During the execution of a DO loop, an iteration will pass through four states, namely, idle, pending, ready, and nished states. An iteration is idle if none of its dependent predecessors have completed; an iteration is pending if some of its dependent predecessors have completed, but not all; an iteration is ready if all its dependent predecessors have completed, but itself has not; otherwise, an iteration is nished. In addition, an iteration without any dependent predecessors is called an initial iteration, which can only have ready and nished states. Via describing an execution scheme, this paper studies the characteristics of Fortran DO loops which are related to the eeciency of the execution. Speciically, this paper investigates (1) the number of initial iterations, (2) the maximum number of ready iterations at any instances during the execution, (3) the maximum number of pending iterations at any instances during the execution, (4) a hash function to disperse diierent pending iterations, and (5) the parallel execution time.