Improving Speculation Accuracy with Inter-thread Fetching Value Prediction

摘要

Conventional software speculative parallel models are facing challenges due to the increasing number of the processor core and the diversification of the application. The speculation accuracy is one of the key factors to the performance of software speculative parallel model. In this paper, we proposed a novel value prediction mechanism named Inter-thread Fetching Value Prediction(IFVP). It supports a speculative thread to read the values of conflict variables speculatively from another speculative thread. This method can remarkably reduce the miss speculation rate in a loop to be parallelized with cross-iter dependencies. We have proved that the IFVP can improve the speculation accuracy by about 19.1% on the average, and can improve the performance by about 37.1% on the average, compared with the conventional models without value prediction.