Design of n-Gram Based Dynamic Pre-fetching for DSM

Abstract

Many earlier reported works have shown that data pre-fetching can be an efficient answer to the well-known memory stalls. If one can reduce these stalls, it leads to performance improvement in terms of overall execution time for a given application. In this paper we propose a new n-gram model for prediction, which is based on dynamic pre-fetcher, in which we compute conditional probabilities of the stride sequences of previous n steps. Here n is an integer which indicates data elements. The strides that are already pre-fetched are preserved so that we can ignore them if the same stride number is referenced by the program due to principle of locality of reference, with the fact that it is available in the memory, hence we need not pre-fetch it. The model also gives the best probable and least probable stride sequences, this information can further be used for dynamic prediction. Experimental results show that the proposed model is far efficient and presents user certain additional input about the behavior of the application. The model flushes once number of miss-predictions exceed pre-determined limit. One can improve the performance of the existing compiler based Software Distributed Shared Memory (SDSM) systems using this model.