Persistent storage for distributed applications

Abstract

Research work in supporting distributed applications has traditionally focussed on the dynamic part of their interactions—that is, network communication paths. It’s our belief that although these are important, their very transience means that they are much less valuable in the long term than the persistent state that these applications manipulate and leave behind. Such state can be enormous—tens of terabytes are not uncommon for large-scale commercial applications, which are frequently constructed from suites of federated, distributed applications. Such systems are themselves classic examples of a distributed application composed from a set of cooperating pieces. As with network communications, the persistent storage medium must be well-behaved, in the sense of providing predictable behavior, so that applications do not interfere with each other. We believe that quality of service (QoS) guarantees, and ways to automatically reason about resource provision to meet them, is the key to building effective and useful storage services.