Evaluation and Analysis of In-Memory Key-Value Systems

This paper presents an in-depth measurement study of in-memory key-value systems. We examine in-memory data placement and processing techniques, including data structures, caching, performance of read/write operations, effects of different in-memory data structures on throughput performance of big data workloads. Based on the analysis of our measurement results, we attempt to answer a number of challenging and yet most frequently asked questions regarding in-memory key-value systems, such as how do in-memory key-value systems respond to the big data workloads, which exceeds the capacity of physical memory or the pre-configured size of in-memory data structures? How do in-memory key value systems maintain persistency and manage the overhead of supporting persistency? why do different in-memory key-value systems show different throughput performance? and what types of overheads are the key performance indicators? We conjecture that this study will benefit both consumers and providers of big data services and help big data system designers and users to make more informed decision on configurations and management of key-value systems and on parameter turning for speeding up the execution of their big data applications.