A pattern language for parallelizing irregular algorithms

Abstract

In irregular algorithms, data set’s dependences and distributions cannot be statically predicted. This class of algorithms tends to organize computations in terms of data locality instead of parallelizing control in multiple threads. Thus, opportunities for exploiting parallelism vary dynamically, according to how the algorithm changes data dependences. As such, effective parallelization of such algorithms requires new approaches that account for that dynamic nature. This dissertation addresses the problem of building efficient parallel implementations of irregular algorithms by proposing to extract, analyze and document patterns of concurrency and parallelism present in the Galois parallelization framework for irregular algorithms. Patterns capture formal representations of a tangible solution to a problem that arises in a well defined context within a specific domain. We document the said patterns in a pattern language, i.e., a set of inter-dependent patterns that compose well-documented template solutions that can be reused whenever a certain problem arises in a well-known context.