Speaker: |
Ugur Cetintemel University of Maryland, College Park |
---|---|
Topic: | Decentralized Data Consistency Protocols for Mobile and Wide-Area Environments |
Date: | Tuesday, March 27, 2001 |
Time: | 11:00 AM |
Place: | Gould-Simpson, Room 701 |
Replication is essential to systems that attempt to efficiently support shared data in mobile and wide-area environments. Traditional approaches for managing replicated data are often ill-suited for these environments because of implicit assumptions of high availability, strong connectivity, and static environmental and application-specific characteristics. This talk will describe two systems that address these concerns by integrating decentralized commitment, peer-to-peer information propagation, and light-weight adaptation mechanisms.
The first part of the talk will cover Deno, an object-replication system specifically designed for use in mobile and weakly-connected environments. At the core of Deno lies a decentralized update commitment protocol that is implemented through peer-to-peer, epidemic information flow. Unlike previous protocols, this combination allows Deno to support strong consistency without sacrificing availability. The second part of the talk will cover my ongoing work on ReBound, a middleware system that supports a general numerical divergence control framework for deploying distributed applications that can tolerate and benefit from bounded inconsistency. ReBound enables clients to specify continuous or one-time (i.e., per-read) quantitative precision constraints on the data they read. ReBound incorporates peer-to-peer, decentralized server-side protocols that efficiently ensure that these constraints are met by refreshing client caches as necessary. In both Deno and ReBound, light-weight adaptation is facilitated through the use of system-wide fixed weights and pair-wise weight redistribution.
The talk will also present experimental results, based on the Deno and ReBound prototypes, that demonstrate the performance advantages of our protocols over existing decentralized approaches.