Building A Scalable And High-Performance Key-Value Store System

Contemporary web sites can store and process very large amounts of data. To provide timely service to their users, they have adopted key-value (KV) stores, which is a simple but effective caching infrastructure atop the conventional databases that store these data, to boost performance. Examples are Facebook, Twitter and Amazon. As yet little is known about the realistic workloads outside of the companies that operate them, this dissertation work provides a detailed workload study on Facebook\u27s Memcached, which is one of the world\u27s largest KV deployment. We analyze the Memcached workload from the perspective of server-side performance, request composition, caching efficacy, and key locality. The observations presented in this dissertation lead to several design insights and new research direction for KV stores - Hippos, a high-throughput, low-latency, and energy-efficient KV-store implementation. Long considered an application that is memory-bound and network-bound, re- cent KV-store implementations on multicore servers grow increasingly CPU-bound instead. This limitation often leads to under-utilization of available bandwidth and poor energy efficiency, as well as long response times under heavy load. To address these issues, Hippos moves the KV-store into the operating system\u27s kernel and thus removes most of the overhead associated with the network stack and system calls. It uses the Netfilter framework to quickly handle UDP packets, removing the overhead of UDP-based GET requests almost entirely. Combined with lock-free multithreaded data access, Hippos removes several performance bottlenecks both internal and external to the KV-store application. Hippos is prototyped as a Linux loadable kernel module and evaluated it against the ubiquitous Memcached using various micro-benchmarks and workloads from Face- book\u27s production systems. The experiments show that Hippos provides some 20- 200% throughput improvements on a 1Gbps network (up to 590% improvement on a 10Gbps network) and 5-20% saving of power compared with Memcached