Investigating the robustness of adaptive Dynamic Loop Scheduling on heterogeneous computing systems

Dynamic Loop Scheduling (DLS) algorithms are a powerful approach towards improving the performance of scientific applications via load balancing. The adaptive DLS (ADLS) methods have been proven to be the most appropriate for effectively balancing such applications, due to the fact that they are designed to address highly irregular, stochastic behavior caused by algorithmic and systemic variations. To guarantee certain performance levels of such DLS methods, metrics are required to measure their robustness against various unpredictable variations of factors in the computing environment. In this paper, the focus is on investigating metrics for the robustness of two Adaptive Weighted Factoring (AWF) techniques, AWFB and AWFC, as well as of the Adaptive Factoring (AF) technique. Two robustness metrics, called flexibility and resilience, are formulated for these techniques. We also discuss their computational complexity and give notes on their usefulness.