|
|
|
|
|
David Yates University of Massachusetts, Amherst Faculty Candidate, Department of Computer Science Date : Tuesday, March 15, 2005 Time : 9:30 - 10:30 a.m. Location : Science & Tech II, Room 430A Abstract: Challenging issues of scale for networked servers arise from the number of clients that may access each server. The first part of this talk describes two complementary methods for scaling data delivery from networked servers to a large number of clients. The first method we describe is connection-level parallelism (where different connections are processed in parallel) for network protocols running on a shared memory multiprocessor server. We show that different implementation choices are appropriate for sending streaming media when compared with sending conventional data, such as web pages and images. For streaming media, aggregate throughput and the manner in which throughput is distributed across connections are both important performance measures. Matching the number of threads to the number of physical processors in the system yields the best overall performance when sending continuous media. Delivering the desired rate to each connection is accomplished by having threads directly measure their own performance and feed this information to the scheduler. Aggregate throughput is the most important performance criteria when sending conventional data. In this case, matching the number of threads to the number of connections in the system yields the best performance. The second method we describe is scaling data delivery using distributed servers. We investigate two important components of distributing and delivering content via distributed servers: (1) ensuring that each server has the most valuable content for the clients it is serving; and (2) ensuring that load is dynamically shared among servers during peaks in workload from clients. These techniques were both developed in a commercial setting and are integral to current network operations at several service providers. In wireless sensor networks, challenging issues of scale arise from the number of sensor nodes that need to deliver data to sensor field gateways. In the final part of this talk we explore the benefits and costs of caching and prefetching for sensor network-based applications. In such systems, one or more driving applications retrieve sensor data that is disseminated from sensor fields. As data is consumed by these applications, they have an associated quality that is a function of two parameters: (1) the delay experienced in receiving the data, and; (2) the divergence between the value of each data item and the corresponding data in the sensor field. Acquiring sensor data also has a cost that represents the system resources consumed in querying and delivering the data to the applications. This work examines how well different data management approaches perform in this quality versus cost trade-off, given that the sensor network should support fixed cost, and both fixed and variable quality. Our work demonstrates that different data management approaches should be used depending on the application requirements and the characteristics of the environment and sensor network. Bio: David graduated with a Bachelors degree in Electrical Engineering and Computer Science from Tufts University in 1989. David earned a Masters degree in Computer Science from the University of Massachusetts at Amherst in 1992. He is currently finishing his PhD at the University of Massachusetts, and expects to graduate in May. David also co-founded InfoLibria in 1997, led software development there, and then sold the company in 2003. InfoLibria began as a spin-off from Boston University, and grew to be a leading supplier of content distribution and delivery products.
Seminar Point of Contact: Prof. Jana Kosecka
Designated as a Center of Academic Excellence in Information Assurance Education by the National Security Agency
The Committee on National Security Systems
and the National Security Agency have certified that George Mason
University offers a set of courseware that has been reviewed by National Level Information Assurance Subject
Matter Experts and determined to meet National Training Standards for Information Systems Security Professionals,
NSTISSI No. 4011,
4012, and
4013
for academic years 2005 - 2008.
|
