Value Predicted LogSPoTM: Improve the Parallesim of Thread Level System by Using a Value Predictor
Bobin Deng, Hong An, Qi Li, Gongming Li, Mengjie Mao · 2012
Thread Level Speculation (TLS) and Transactional Memory (TM) have been proposed and developed to solve the difficulty of parallel programming during the last decades. Even though researchers have make effort to optimize these two kinds of systems for a long time, TLS and TM still have not been largely applied in commercial fields due to their wobbly performance. As we know, these two systems will show attractive performance improvement while they are running the applications with low data contentions. However, when facing some high confliction programs, the system performance will become awfully. Just in order to find a way out of this difficult situation, we raise a value predictor scheme in this paper. We implement our value predictor on LogSPoTM, which is a hybrid TLS and TM system with moderate hardware complexity. From our simulation results of 8 benchmarks, we can see that with the help of the value predictor, this new LogSPoTM system achieves 64.8% performance improvement in best case and about 30% on average, which may have the potentiality to enlarge application fields of TLS/TM system extensively.