Scheduling on the Grid via multi-state resource availability prediction

To make the most effective application placement decisions on volatile large-scale heterogeneous Grids, schedulers must consider factors such as resource speed, load, and reliability. Including reliability requires availability predictors, which consider different periods of resource history, and use various strategies to make predictions about resource behavior. Prediction accuracy significantly affects the quality of the schedule, as does the method by which schedulers combine various factors, including the weight given to predicted availability, speed, load, and more. This paper explores the question of how to consider predicted availability to improve scheduling, concentrating on multi-state availability predictors. We propose and study several classes of schedulers, and a method for combining factors. We characterize the inherent tradeoff between application makespan and the number of evictions due to failure, and demonstrate how our schedulers can navigate this tradeoff under various scenarios. We vary application load and length, and the percentage of jobs that are checkpointable. Our results show that the only other multi-state prediction based scheduler causes up to 51% more evicted jobs while simultaneously increasing average job makespan by 18% when compared with our scheduler.