Reordering Iterations in Runtime Loop Parallelization Reordering Iterations in Runtime Loop Parallelization

When a loop in a sequential program is parallelized, it is normally guaranteed that all ow dependencies and anti-dependencies are respected so that the result of parallel execution is always the same as sequential execution. In some cases, however, the algorithm implemented by the loop allows the iterations to be executed in a di erent sequential order than the one speci ed in the program. This opportunity can be exploited to expose parallelism that exists in the algorithm but is obscured by its sequential program implementation. In this paper, we show how parallelization of this kind of loop can be integrated into the runtime parallelization scheme of Saltz et al. [17, 18]. Runtime parallelization is a general technique appropriate for loops whose dependency structures cannot be determined at compile time. The compiler generates two pieces of code: the inspector examines dependencies at run time and computes a parallel schedule; the executor executes iterations in parallel according to the computed schedule. In our case, the inspector has to solve two problems: choosing an appropriate sequential order for the iterations and computing a parallel schedule. The two problems are treated as a single graph coloring problem, which is solved heuristically. Two methods to do so are described. Furthermore, the basic runtime parallelization scheme for shared-memory multiprocessors pays no attention to locality when scheduling iterations onto processors. One of our methods takes locality consideration into account when making these decisions. The performance implications of reordering are examined experimentally on a KSR1 parallel machine as well as through a simple analytic model of execution time.