Closed-loop Adaptive and Stochastic Prefetch Mechanism for Data Array

Projet ANR PersyvalInternational audienceNowadays, one of the main limiting factor in multi-processordevelopment is the increasing speed gap between efficient processingelements and slow main memories. To reduce this limitation, prefetchingmechanisms, implemented in memory hierarchy, attempt to predict thefuture data needed in local memory. However, classical proposedsolutions are no longer efficient for typical access sequences in thecontext of image processing algorithms.In this paper, a prefetching mechanism is proposed, based on the use ofthe index array of access rather than memory addresses. The adaptiveprediction policy takes into account both application and main memorycharacteristics. This mechanism enhances the Cache nD-AP, formerlystatically set, to improve its performance. Experiments show that theproposed adaptive strategy is able to maintain its efficiency when thecharacteristics of the system dynamically change