Exploiting memory locality in distributed key-value stores

Distributed key-value stores which have a main-memory-based persistence can be utilized for serving applications whose performance criteria cannot be met by scaling-out across disk-based nodes. This paper outlines a PhD thesis that addresses the resulting challenge of exploiting memory locality in such systems. We propose to introduce the notion of a block that allows to store a certain number of values consecutively in main-memory. Preliminary micro-benchmarks indicate that this could increase the throughput by a factor 10 on a per-node level. Furthermore, this paper presents a research agenda that covers the resulting implications on a per-node and system-level.