Computer Science Department

Computer Science Colloquium
bar

TACT: Continuous Consistency for Wide-Area Replication

Haifeng Yu
Duke University

Friday, April 4, 2003
11:30 a.m.
Room 1302 WWH
251 Mercer Street
New York, NY 10012-1185

Host: Richard Cole, cole@cs.nyu.edu, 212-998-3119
Directions: http://cs.nyu.edu/csweb/Location/directions.html
Colloquium Information: http://cs.nyu.edu/csweb/Calendar/colloquium/index.html

Abstract

Replication is a key approach for scaling wide-area applications and achieving high-performance and high-availability. However, because of wide-area latency and potential Internet congestion/failures, consistency overheads can easily offset the benefits of replication. This duality currently limits the deployment of replicated services across the wide area.

To address such problem, in this talk I will explore the spectrum between the traditional two extremes of {\em strong consistency} and {\em optimistic consistency}, and propose the concept of {\em continuous consistency}. Continuous consistency overcomes the fundamental obstacle in wide-area replication by allowing applications to choose middle points on the consistency spectrum. Specifically, I will describe the design of the TACT middleware that enables dynamically tunable consistency levels. I will also discuss the TACT implementation and performance evaluation across the Internet. The results show that applications can achieve significant semantics and performance benefits using the TACT middleware.

Next, I will explore the consistency model's effects on system availability, both theoretically and experimentally. I will first discuss a tight availability upper bound for all consistency protocols. Then I will describe how I measure the availability of existing consistency protocols in both wide-area deployment and a local emulation environment, and compare the results to the upper bound. I will show that: i) simple optimizations can significantly improve availability; and ii) for small-scale systems, the optimized protocols can approach the upper bound in my experiments.

Biography

Haifeng Yu is a Visiting Assistant Professor of Computer Science Department, Duke University. His research interests cover the general area of distributed systems, as well as related fields such as operating systems, database systems, fault-tolerance and large-scale peer-to-peer systems. Haifeng receives his Ph.D. and M.S. from Duke University, and his B.E. from Shanghai Jiaotong University, China. More information about his research is available at http://www.cs.duke.edu/~yhf

bar
e-mail: webmaster@cs.nyu.edu