Power and Performance Analysis of Persistent Key-Value Stores

TL;DR

This paper compares the power and performance of RocksDB and Kreon key-value stores across different server architectures, revealing microservers' superior power efficiency with similar latency.

Contribution

It provides a comparative analysis of persistent key-value stores on diverse hardware architectures, highlighting power efficiency benefits of microservers.

Findings

Microservers are 0.68-3.6x more power-efficient.

Kreon reduces CPU overhead compared to RocksDB.

Power efficiency gains are consistent across server generations.

Abstract

With the current rate of data growth, processing needs are becoming difficult to fulfill due to CPU power and energy limitations. Data serving systems and especially persistent key-value stores have become a substantial part of data processing stacks in the data center, providing access to massive amounts of data for applications and services. Key-value stores exhibit high CPU and I/O overheads because of their constant need to reorganize data on the devices. In this paper, we examine the efficiency of two key-value stores on four servers of different generations and with different CPU architectures. We use RocksDB, a key-value that is deployed widely, e.g. in Facebook, and Kreon, a research key-value store that has been designed to reduce CPU overhead. We evaluate their behavior and overheads on an ARM-based microserver and three different generations of x86 servers. Our findings show that microservers have better power efficiency in the range of 0.68-3.6x with a comparable tail latency.