Scheduling-based contention management techniques for transactional memory

Contention management refers to the mechanisms used by transactional memory (TM) implementations “to ensure forward progress – to avoid livelock and starvation, and to promote throughput and fairness” [1]. Without effective contention management mechanisms, TM implementations are susceptible to performance degradation caused by numerous transaction collisions.Early work on contention management focused on the narrower problem of conflict resolution. When two transactions collide, one transaction (the winner transaction) is allowed to proceed, while the other (the loser transaction) must wait and/or be aborted. Conflict resolution policies decide which transaction should win and which should lose and for how long the losing transaction should be delayed. However, it was shown that conflict resolution alone is insufficient for guaranteeing reasonable performance for high-contention TM workloads.The key idea underlying transaction schedulers, introduced a few years ago, is that the execution of conflicting transactions must be serialized in the face of high contention and, more generally, that the level of parallelism between transactional threads should be controlled by the contention manager and dynamically adjusted. Transaction scheduling allows not only to resolve conflicts after they occur, but also to proactively reduce their probability, thus improving performance. This chapter provides a survey of the key approaches and techniques used by transaction schedulers.