Scheduling Transactions in Replicated Distributed Software Transactional Memory

Distributed software transactional memory (DTM) is an emerging, alternative concurrency control model for distributed systems that promises to alleviate the difficulties of lock-based distributed synchronization. Object replication can improve concurrency and achieve fault-tolerance in DTM, but may incur high communication overhead (in metric-space networks) to ensure one-copy serializability. We consider metric-space networks and develop a cluster-based object replication model for DTM. In this model, object replicas are distributed to clusters of nodes, where clusters are determined based on distance between nodes, to maximize locality and fault-tolerance and to minimize communication overhead. We develop a transactional scheduler for this model, called CTS. CTS enqueues live transactions and identifies some of the transactions that must be aborted in advance to enhance concurrency of the other transactions over clusters, reducing a significant number of future conflicts. Our implementation and experimental evaluation reveals that CTS improves transactional throughput over state-of-the-art replicated DTM solutions by as much as (average) 1.55x and 1.73x under low and high contention, respectively.