DATE: Wednesday, September 28, 2005 Note Special Day
TIME: 3:30 pm - 4:30 pm Note Special Time

Mike Dahlin
Univ. of Texas at Austin

Towards a Unified Theory and Practice of Large-Scale Replication

This talk argues for developing a "unified theory and practice" for constructing large-scale replication systems that span large networks or mobile nodes. It then describes PRACTI, a first step towards this goal. PRACTI is the first replication protocol to simultaneously provide three key properties: (1) Partial Replication of data so that each node only needs store and to see updates for data of interest, (2) Arbitrary Consistency so that different applications can enforce the consistency requirements they need, and (3) Topology Independence so that data propagation among nodes can be optimized.

A unified theory of replication would yield two significant benefits. First, providing a set of the right core abstractions would dramatically simplify the design and deployment of new systems. Second, cleanly separatng mechanism from policy would provide better performance/availability/consistency trade-offs than are now available to existing "point-solution" systems.

Our initial experience with PRACTI suggests that both benefits may be significant. The PRACTI mechanisms subsume a broad range of existing replication architectures and have allowed us to quickly construct a number of interesting replication systems. And, not only are replication systems easier to build under PRACTI, we have often been able to make dramatically better performance trade-offs than made by existing systems: for workloads of interest, our PRACTI design dominates existing approaches by providing an order of magnitude better bandwidth and storage efficiency than replicated server systems (AC-TI), an order of magnitude better synchronization cost compared to hierarchical systems (PR-AC), and consistency guarantees not achievable by per-object replication systems (PR-TI).

Mike Dahlin is an Associate Professor and Faculty Fellow in the Department of Computer Sciences at the University of Texas at Austin. His work focuses on large scale distributed systems. Dr. Dahlin received his PhD from the University of California at Berkeley in 1996, the NSF CAREER award in 1998, and the Sloan Research Fellowship in 2000.

HOST: Haifeng Yu


For Further Seminar Info Contact:
or visit