Genuine replication, opacity and wait-free read transactions: can a STM get them all?

Distributed Transactional Memory (DTM) is an attractive programming paradigm that aims at combining the simplicity of Transactional Memory (TM) with the scalability and failure resiliency achievable by exploiting the resource redundancy of distributed platforms. In these in-memory distributed platforms replication plays a role of paramount importance both for fault-tolerance and performance reasons. The advantages of replication from the fault-tolerance’s perspective are quite obvious. From the performance’s standpoint, however, replication can become a double-sided weapon. By maintaining copies of the same datum across multiple nodes, in fact, the efficiency of a DTM platform can be enhanced by allowing more than one node to access that datum locally (sparing the costs of remote RPCs). On the other hand, replication has an inherent cost as it demands additional, costly operations to enforce synchronization among replicas. Hence, if the costs of replication are not outweighed by the benefits it brings, replication may ultimately hinder performance rather than boosting it. In this talk we focus on the problem of designing efficient replication protocol for large scale DTM platforms. We argue that efficient replication protocols for large scale DTM systems should follow these three key design principles: