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.  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 for a broad range of network services.  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.

After presenting the motivation and design of TACT, I will describe the results of a number of experiments quantifying the tradeoff between performance, availability, and consistency for a number of sample Internet services running across the wide area.  Next, I present a technique for calculating tight upper bounds on the availability of replicated services as a function of workload, fault load, and desired level of consistency and compare the achieved availability of replicated services using existing consistency protocols relative to this upper bound.  Simple optimizations to these protocols allow services to approach our calculated upper bound for a number of measured faultloads.

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

BIOGRAPHY: