Arbiter work stealing for parallelizing games on heterogeneous computing environments
Wessam AlBahnassi, Dhrubajyoti Goswami, Sudhir P. Mudur · High Performance Computing Symposium · 2013
Games are simulations of the physical and imaginary worlds. Games nowadays run on commodity platforms that include different categories of powerful computing elements with varying capabilities. To benefit from this variety, suitable mapping of works to computing elements is essential for optimal performance. Arbiter Work Stealing (AWS) is a new scheduler addressing this requirement. The AWS scheduler builds on the classical work stealing algorithm by adding an upper layer that manages multiple running instances of the work stealing algorithm. AWS automatically schedules the dynamically generated game application tasks to appropriate processors using a cost model that takes into account current work load, execution times, data locality, and data transfer rates. Experimental results show that incorporating AWS to schedule tasks of a parallel game application yields superior performance through better utilization of the available resources and through better use of data locality in a heterogeneous computing environment.