High Performance Design for Redis with Fast Event-Driven RDMA RPCs

Autor: Chengcheng Huang, Huiqi Hu, Xing Wei, Aoying Zhou, Xuan Zhou, Xuecheng Qi
Rok vydání: 2020
Předmět:
Zdroj: Database Systems for Advanced Applications ISBN: 9783030594091
DASFAA (1)
DOI: 10.1007/978-3-030-59410-7_12
Popis: Redis is a popular key-value store built upon socket interface that remains heavy memory copy overhead within the kernel and considerable CPU overhead to maintain socket connections. The adoption of Remote Direct Memory Access (RDMA) that incorporates outstanding features such as low-latency, high-throughput, and CPU-bypass make it practical to solve the issues. However, RDMA is not readily suitable for integrating into existing key-value stores. It has a low-level programming abstraction and the original design of existing systems is a hurdle to exploit RDMA’s performance benefits. RPCs can provide simple abstract programming interfaces that make it easy to be integrated into existing systems. This paper proposes a fast event-driven RDMA RPC framework named FeRR to promote the performance of Redis. First, we describe our design of FeRR that is based on one-sided RDMA verbs. Second, we propose an efficient request notification mechanism using event-driven model that can decrease the CPU consumption of polling requests. Finally, we introduce a parallel task engine to eschew the bottleneck of the single-threaded execution framework in Redis. Comprehensive experiment shows that our design achieves orders-of-magnitude better throughput than Redis - up to 2.78 million operations per second and ultra-low latency - down to 10 \(\upmu \)s per operator on a single machine.
Databáze: OpenAIRE