DDASTM: Ensuring Conflict Serializability Efficiently in Distributed STM

TLDR: This paper proposes an approach to make each cycle in PG detected by those transactions, which construct this cycle, together in parallel way, instead of detecting cycle individually, and shows that the average execution time and communication cost of all transactions can be decreased.

Abstract: CS (Conflict Serializability) is a recently proposed relaxer correctness criterion that can increase transactional memory’s parallelism. DDA (Distributed Dependency-Aware) model is currently proposed to implement CS in distributed STM (Software Transactional Memory) for the first time. However, its transactions detect conflicts individually via detecting cycles in PG (Precedence Graph) and cause extra runtime overhead, especially at the condition that the transactions access lots of objects or the PG is large. In this paper, we propose an approach to make each cycle in PG detected by those transactions, which construct this cycle, together in parallel way, instead of detecting cycle individually. Experimental results show that the average execution time and communication cost of all transactions, including aborted ones, in our approach, can be decreased to 76% and 78% of those in DDA respectively. Its speedup is up to 2.56× against baseDSTM, employing two-phase locking.