COLLOQUIUM
             Computer Science Department, Boston University

Speaker: Magdalena Balazinska
         MIT

Date:  Monday, March 21
Time:  9:00
Place: Room MCS 135, 111 Cummington Street 
       (for directions, see www.cs.bu.edu/colloquium)

Title: Load Management and Fault-Tolerance in a 
       Distributed Stream Processing System

Abstract:

Recently, a new class of data management applications has emerged in
areas such as sensor-based environmental monitoring, financial
services, network monitoring, and military applications. These "stream
processing applications" require low-latency processing of
large-volume data streams. Because traditional database management
systems are ill-suited for high-volume, low-latency stream processing,
new systems, called stream processing engines (SPEs), have been
developed.

In this talk, we present the software architecture and algorithms in
Borealis, one of the first distributed stream processing engines. We
discuss how our system meets two important challenges: (1) distributed
load management, and (2), fault-tolerant operation in the face of node
failures, network failures, and network partitions.

We present a mechanism that enables autonomous participants to
collaboratively handle load. Our approach is based on contracts that
participants negotiate offline. At runtime, participants move load
only to partners with whom they have a contract and pay each other the
contracted price, making the mechanism lightweight. We show that our
approach provides incentives that foster participation and leads to
good system-wide load balance properties.

For fault-tolerance, we present a replication-based scheme that masks
most node and network failures. When network partitions occur, our
approach addresses the traditional availability-consistency trade-off
by striving to minimize inconsistencies, while ensuring that the
system meets the desired availability specified by the application or
user.

Biography:

Magdalena Balazinska is a PhD candidate in the Networks and Mobile
Systems Group at the MIT Computer Science and Artificial Intelligence
Laboratory. Her research interests focus on distributed data
management systems. She is one of the major contributors to
Borealis/Medusa, a distributed stream processing engine developed at
MIT, Brown University, and Brandeis University. As part of her
dissertation work, she is devising, implementing, and evaluating
various load management and fault-tolerance techniques for distributed
and federated stream processing. During the course of her PhD,
Magdalena has also worked on various other data management problems
including scalable resource discovery, nomadic data-access, and
censorship circumvention. Before joining MIT, Magdalena received a
B.E. and M.S. from Ecole Polytechnique de Montreal.

Host: George Kollios