E E C S  MIT Electrical Engineering and Computer Science

EECS Event

Building Replicated Internet Services Using a Continuous Consistency Model

Amin Vahdat
Duke University

Tuesday, February 27, 2001
2:00 PM (refreshments 1:45)
Room NE43-518
POCS (Principles of Computing)

Abstract

System designers wishing to build fast and highly available Internet services must contend with network congestion, latency, and unpredictable failures. Replication is a key approach for improving service performance and availability. Unfortunately, the benefits of replication are limited by the overhead of maintaining consistency across the wide area. The overhead of strong consistency protocols limits practical wide-area replication to a handful of sites. While many Internet services do not require strong consistency, existing optimistic consistency models allow replicas to become arbitrarily stale.

This talk describes TACT, a continuous consistency model that allows replicas to bound their divergence from strong consistency. A spanning set of metrics---Numerical Error, Order Error, and Staleness---captures the consistency semantics of a broad range of network services. TACT enforces consistency at the granularity of application-specific "conits" (consistency units). Using these metrics, applications can dynamically trade consistency for performance and availability based on changing client, network, and service characteristics. Fundamentally, we aim to determine upper bounds on the performance and availability of replicated services as a function of workload, faultload, and consistency.

An initial wide-area evaluation of three sample services using TACT confirms the utility of the model in trading consistency for performance. Currently, we are evaluating the availability achieved by this prototype relative to upper bounds calculated analytically. We are also using the idea of bounded inconsistency for metadata dissemination in wide-area networks. Content distribution networks, active networks, web hosting services, and virtual supercomputers must track the status of replicas to perform global functions such as request routing, resource allocation, or load balancing. We are developing a communication abstraction, dicast, that allows replicas to specify the rate of information decay as a function of distance from the local node.

More information is available at http://www.cs.duke.edu/~vahdat

Hosts: Hari Balakrishnan and Robert Morris


Related pages: 2000-01 events   |   Current events   |   2000-01 Web archives
This page:
http://www-eecs.mit.edu/AY00-01/events/44.html
Created: Feb 26, 2001  |  Modified: Feb 26, 2001
Site table of contents  |  Site map  |  Search  |  Your comments and inquiries are welcome.