Esodyp+: Prefetching in the Jackal Software DSM

Prefetching transfers a data item in advance from its storage location to its usage location so that communication is hidden and does not delay computation. We present a novel prefetching technique for object-based Distributed Shared Memory (DSM) systems and discuss its implementation. In contrast to page-based DSMs, an object-based DSM distributes data on the level of objects, rendering current prefetchers for page-based DSMs unsuitable due to more complex data streams. To predict future data accesses, our prefetcher uses a new predictor (Esodyp+) based on a modified Markov model that automatically adapts to program behavior. We compare our prefetching strategy with both a stride prefetcher and the prefetcher of the Delphi DSM system. For several benchmarks our prefetching strategy reduces the number of network messages by about 60%. On 8 nodes, runtime is reduced by 15% on average. Hence, network-bound programs benefit from our solution. In contrast to the other predictors, Esodyp+ achieves a prediction accuracy above 80% with only 8% of unused prefetches for the benchmarks.