Addressing the idiosyncrasies of context-dependent parallelization

Abstract

Parallelization of sequential code to yield an increase in the execution time on a multiprocessor system is a mature field of research. Patterns or dwarfs that frequently occur among the sequentially coded grand challenge applications have been comprehensively identified and categorized in literature. POVRay, a ray tracing based grand challenge application, that also belongs to the SPEC suite of benchmarks, demonstrates a unique as yet unclassified pattern that may place it in a separate class of applications. The execution profile of POV-Ray is exceedingly dependent on an external file supplied as input. Conventional parallelization methods do not consider the impact that an external input file may have on the execution profile of the application. In the case of POV-Ray, a parallelized executable may result in a speedup less than one when used with different input files as a result of the increased communication-to-cost ratio. This paper elaborates on this unique feature of POV-Ray and presents an argument as to why the conventional method of parallelizing will not work in this case. A novel context-dependent parallelization is suggested that uses scaled-down profiling to generate context-dependent executables to ensure a speedup greater than one in all cases.