Self-Organizing Subsets: From Each According to His Abilities, to Each According to His Needs

The key principles behind current peer-to-peer research include fully distributing service functionality among all nodes participating in the system and routing individual requests based on a small amount of locally maintained state. The goals extend much further than just improving raw system performance: such systems must survive massive concurrent failures, denial of service attacks, etc. These efforts are uncovering fundamental issues in the design and deployment of distributed services. However, the work ignores a number of practical issues with the deployment of general peer-to-peer systems, including i) the overhead of maintaining consistency among peers replicating mutable data and ii) the resource waste incurred by the replication necessary to counteract the loss in locality that results from random content distribution. This position paper argues that the key challenge in peer-to-peer research is not to distribute service functions among all participants, but rather to distribute functions to meet target levels of availability, survivability, and performance. In many cases, only a subset of participating hosts should take on server roles. The benefit of peerto-peer architectures then comes from massive diversity rather than massive decentralization: with high probability, there is always some node available to provide the required functionality should the need arise. This research is supported in part by the National Science Foundation (EIA-9972879, ITR-0082912), Hewlett-Packard, IBM, Intel, and Microsoft. Braynard and Reynolds are supported by an NSF graduate fellowships and Vahdat is also supported by an NSF CAREER award (CCR-9984328).