Two-Phase Conflict Detection for Transactional Memory on Clusters

Ruibo Wang, Kai Lu, and Xicheng Lu
Computer School, National University of Defense Technology, Changsha, China

Transactional Memory (TM) research has focused on multi-core processors; limited research has been aimed at the clusters. The intention of deploying TM on clusters is using more processors to solve big problems with this convenient technique. But the performance of the existing cluster's TM is very poor. The conflict detection, which is the most frequent operation of TM, is highly depending on the memory access. But the cluster's remote memory access speed is very slow. The TM design derived from the multi-core processor is not suitable for the clusters.

Now we introduce a hierarchical conflict detection strategy. The detection process was divided into two phases, "inside a node" and "among the nodes". By using different detection method with the two-phase, the hierarchical strategy provides a good practical performance. Our early results indicate the two-phase strategy shows better performance and scalability when the concurrent transaction count is greater than the node number of the cluster. This paper motivates more detailed study of using TM on clusters.

Long Abstract

Back to Program