Optimized Memcached on User-Space TCP Stack

Abstract

Memcached is a wildly used multi-thread, distributed Key-Value caching solution in todays many web applications and services to provide high query performance and reduce tail latency. Optimizing memcached has always been a hot topic in both academia and industry from various perspective. In this paper, we describe a solution for memcached that utilizes user space TCP stack on top of intel DPDK to improve its performance on multi-core system. This solution enables the system to process key-value queries bypassing the linux kernel to greatly reduce network overhead especially in situation where the system needs to exchange a huge amount of small messages with clients. We implemented the system based on the best optimized version of memcached, MemC3, and used mTCP with Intel DPDK as the user-space TCP stack. Our experimental results showed that MemC3 and DPDK can be greatly combined, and the throughput of the system can be increased by 50% even when it scales to multiple cores. Apart from this, we can also see the tail latency can be significantly reduced by our solution.