To reduce remote cache transfer latency and improve the performance of shared memory systems, a new last-write-touch prediction scheme that exploits the inherent write characteristics of a program is proposed and a directory protocol to support the scheme is adapted. By predicting a last-write-touch and self downgrading a cache block in advance, a processor can get the data from the memory directly and one network hop can be saved for a remote cache access. Compared with the existing instruction-based prediction technique, much storage overhead can be reduced. Experimental results show that it can achieve an average prediction accuracy of 83.1%, leading to improvements up to average 8.57% on the final application performance. Moreover, compared with the instruction-based prediction scheme, the scheme can reduce the storage overheads of the history table by 69% and the storage overheads of the signature table by 36%. ©, 2015, National University of Defense Technology. All right reserved.
