Synchronization phases (to speed up transactional memory)

Abstract

We introduce the Sync-Phase technique for software transactional memory. Time is divided into two different alternating phases: computation phases and commit (synchronization) phases. Transactions are only allowed to commit during a commit phase, which allows execution without validation (or the need for cache synchronization) during computation phases. Phases can be determined through time or other means. The technique scales well and it is applicable, whenever a single commit of a transaction triggers certain tasks, e.g. validations of a read set or extending the validity of objects for time-based transactional memory. For evaluation we implemented the Sync-Phase STM technique for DSTM and compared it to other techniques implemented based on DSTM, i.e. the original version and one that uses a commit counter to reduce the number of validations. Benchmarks show that Sync-Phase STM allows much faster execution of transactions in most scenarios and scales better than existing approaches relying on a single commit counter.