Slicing as a Distributed Systems Primitive

Abstract

Large-scale distributed systems appear as the major infrastructures for supporting planet-scale services. These systems call for appropriate management mechanisms and protocols. Slicing is an example of an autonomous, fully decentralized protocol suitable for large-scale environments. It aims at organizing the system into groups of nodes, called slices, according to an application-specific criteria where the size of each slice is relative to the size of the full system. This allows assigning a certain fraction of nodes to different task, according to their capabilities. Although useful, current slicing techniques lack some features of considerable practical importance. This paper proposes a slicing protocol, that builds on existing solutions, and addresses some of their frailties. We present novel solutions to deal with non-uniform slices and to perform online and dynamic slices schema reconfiguration. Moreover, we describe how to provision a slice-local Peer Sampling Service for upper protocol layers and how to enhance slicing protocols with the capability of slicing over more than one attribute. Slicing is presented as a complete, dependable and integrated distributed systems primitive for large-scale systems.