Design And Implementation Of Distributed Transaction System Using RDMA And HTM

Transaction,as a powerful programming abstraction,greatly simplifies the relational data processing.With the increase of data size and the level of concurrency,many scenarios,such as web services,stock exchange,and e-commerce,are increasingly required to support low-latency and high-throughput transaction processing.However,traditional transaction processing systems are usually inefficient,fail to support massive data and large-scale concurrent access.At the same time,as the rapid evolution in hardware,many advanced hardware features such as high-capacity(non-volatile)memory,hardware transactional memory(HTM)and fast interconnect with RDMA are becoming available to the market.These advanced hardware technologies offer new opportunities for low-latency,high-throughput transactions: large memory volume makes in-memory transactions feasible,which can significantly reduces I/O overhead in transaction processing;the hardware support for atomicity,consistency,and isolation(ACI)properties makes it very promising to offload concurrency control to CPU,ensuring concurrent performance while significantly reducing engineering complexity;and the RDMA feature further enables fast distributed transactions within a local cluster,which can greatly reduce the latency of distributed transactions.However,while the exising research results have demonstrated the feasibility of combining these advanced hardware features for fast distributed transactions,they fall short in several important aspects.For example,some systems only use a portion of the advanced hardware features,either do not provide distributed transaction support,or rely on complex fine-grained locking mechanisms to achieve transactions;others place several restrictions on transaction such as knowing read/write sets in advance and providing no availability support,or both.This theies presents DrTM+R,a fast and general distributed transaction processing system.Like prior systems,DrTM+R supports in-memory transactions by leveraging battery-backed memory as the main storage for database records and combines HTM and RDMA for fast distributed transactions.Unlike prior systems,DrTM+R places no restrictions on transactional workloads and provides a more general transaction support.DrTM+R addresses the generality issue by designing a hybrid OCC and locking scheme,which leverages the strong atomicity of HTM and the strong consistency of RDMA to preserve strict serializability with high performance.Evaluation using typical OLTP workloads like TPC-C and SmallBank shows that DrTM+R scales well on a 6-node cluster and achieves over 5.69 and 94 million transactions per second for TPC-C and SmallBank respectively,which is an order-of-magnitude faster than a state-of-the-art distributed transaction system(Calvin).