Priority-based squash reducing methods in thread level speculation

Abstract

Thread level speculation (TLS) aggressively transforms long serial programme into multiple short parallel threads to significantly boost the performance of sequential programmes. But frequent squashing, which is caused by violation between multiple parallel threads, will greatly offset the benefits from parallelisation. Most existing works only focus on improving TLS scheme itself directly, such as preventing false-sharing, pre-computing or predicting values and so on. On the contrast, we realise that squashes are caused by messages whose arriving orders are violated. Thus, these squashes can be reduced by rearranging TLS messages. For reducing TLS squashes, in this paper, we first propose a priority-aware network-on-chip (NoC), which uses a priority-based packet arbitration policy to reorder messages at router. Further, we extend this priority scheme by employing prioritising policy into the directory for TLS system that uses directory-based cache coherence protocol. The extension results in a cost-less version. Experimental evaluation for five typical application kernels of SPEC2000 shows that our NoC approach reduces squashing rate by 22% in best case and 15% on average.