Accelerating read atomic multi-partition transaction with remote direct memory access

Many applications these days require data processing that is both efficient and reliable. Distributed databases are one way to meet these requirements, but must be updated using distributed transactions. To manage foreign key constraints, secondary indices, and materialized views in distributed environments, read atomic multi-partition (RAMP) transactions demonstrate high efficiency. RAMP transactions achieved high performance distributed transaction processing by relaxing the isolation level. However, the use of fast interconnect to accelerate performance has not yet been considered. This paper proposes the acceleration of RAMP transactions by exploiting remote direct memory access (RDMA) on the InfiniBand interconnect. We first present GET+ and PUT+ operations that accelerate the RAMP transaction by exploiting RDMA write operations. We then present the GET* operation, which further accelerates GET+ operations exploiting RDMA read operations. To evaluate the proposed methods, we implemented a prototype distributed in-memory key-value store in C/C++. The results of the experiments show that compared with RAMP transactions on IP over InfiniBand, GET* and PUT+ achieve a 2.67× performance improvement on the Yahoo! Cloud Serving Benchmark. All of our code is publicly available.