Conflict Avoidance Scheduling Using Grouping List for Transactional Memory

Abstract

Conventional Transactional Memory (TM) systems may experience performance degradation in applications with high contention, given the fact that execution of transaction will frequently restart due to conflicts. The restarting of transaction essentially requires rollback that is a wasteful operation. To address this point, we developed a system to reduce the overhead caused by high contention. In this paper, we present a method called Conflict Avoidance Scheduling (CAS), which prevents the conflicts in high contention by use of conflict characteristic. In CAS, threads that execute transactions which have high probability of conflicts are grouped together. Based on the group information, concurrent execution of threads in the same group is restricted. Therefore, threads that may cause conflict are serially executed. We evaluate the performance of the proposed design by comparing it with Log TM-SE. The simulation results show that our system improves the performance by 23% on an average in applications with high contention, as compared with the conventional Log TM-SE.