CS Colloquium on Monday, Nov 24 at 11AM

Title: DHTs as Building Blocks for Systems Great and Small:
       Data-Centric Storage for Sensornets and
       OpenHash, a Public DHT Service for the Internet

Speaker: Brad Karp (bkarp@cs.cmu.edu)
         Intel Research and Carnegie Mellon University

Place: MCS 135, 111 Cummington Street

Abstract:

The network research community has devoted significant attention to
Distributed Hash Tables (DHTs) in the past three years: both to
algorithmic primitives with which to build efficient DHTs, and to
diverse complex Internet applications atop DHTs. In this talk, I will
describe reusable DHT-based building blocks for scalable network
systems in two very different contexts: in wireless sensornets, and
in the wired Internet.

In sensornets, the limited battery power of sensor nodes dictates that
delivery of sensed data in response to users' queries must be
energy-efficient. Our novel approach to this data dissemination
problem, Data-Centric Storage (DCS), stores sensed data at a sensor
node determined by the name associated with the sensed data. After
giving a simple model for the energy costs of DCS and two other
canonical approaches to data dissemination, I will describe GHT, a
Geographic Hash Table system for DCS on sensornets. Both the energy
cost model and simulations of GHT demonstrate that DCS disseminates
data more efficiently than the other two canonical approaches.

In the Internet, the vast majority of DHT-based application building
is still done by a small community of DHT researchers. Significant
testbed resources are required to deploy and test DHT-based
applications. Only a tiny fraction of would-be developers has access
to a testbed infrastructure like PlanetLab. If DHTs are to have a
positive impact on the design of broadly used distributed
applications, the community of DHT-based application developers should
be as broad as possible. To spur the broadening of this community, we
propose OpenHash, an open, publicly accessible DHT service that runs
on a set of infrastructure hosts and allows any application to execute
put() and get() operations. Some DHT-based applications require
application-specific processing and thus cannot be limited to a
generic put()/get() interface. We seek to share a common DHT routing
deployment but to allow application-specific functionality to reside
only on certain hosts. OpenHash removes the need in today's DHT
systems to place application-specific code together with DHT code.

[This is joint work with Deborah Estrin, Ramesh Govindan,
 Sylvia Ratnasamy, Scott Shenker, Li Yin, and Fang Yu.]

Bio of the Speaker:

Brad Karp earned a B.S. at Yale University in 1992, an S.M. at Harvard
University in 1995, and Ph.D. at Harvard University in 2000, all in
Computer Science. He is a Staff Researcher at Intel Research Pittsburgh and
an Adjunct Assistant Professor of Computer Science at Carnegie Mellon
University. Prior to joining Intel Research and CMU, he was a researcher
for three years at ICIR, the ICSI Center for Internet Research (previously
named ACIRI) at Berkeley. His research interests include algorithms and
systems for routing, congestion control, and distributed storage in sensor
networks, multi-hop wireless networks, and the Internet.

Host: John Byers (http://www.cs.bu.edu/~byers)