The Colloquium Series of the
            Department of Computer Science,
		  University of Wyoming

			presents

                    Patricia Gilfeather
                 University of New Mexico       

"Making Commodity Protocols Viable in High-Performance Computing"

                  Thursday, March 10, 2005
	                ENG 4066
	             4:00 - 5:00 p.m.

Abstract:

Resource usage in all its forms is the key to systems development.
This talk explores resource constraints in the space between
high-performance computing and commodity computing.  We want to
increase the performance of commodity interconnects for high-end
applications without increasing their cost.  One way to alleviate
performance bottlenecks of commodity networks is to offload parts of
the protocol onto a network interface card (NIC).  However, if
offloading is to be effective, the NIC must remain inexpensive.
Therefore we must address offloading in terms of the CPU and memory
constraints of a commodity NIC.

First, we review the Extensible Message-oriented Offload model (EMO)
designed to study the CPU constraints of offloading processing for
message-oriented, latency-sensitive high-performance applications.  We
illustrate mapping EMO onto the other offload models (for streaming
data) and we model several offloading methods including our Splintered
TCP.  Second, we describe the working set model for open TCP
connections, connection-less TCP, that drastically decreases the
memory footprint for connections without incurring the latency cost of
re-opening the connection.  Finally, we explore how our new models of
offloading commodity protocols for high-performance computing can be
adapted to be competitive in the commodity networking market.

Bio:

Ms. Gilfeather will be defending her Ph.D. in computer science from
the University of New Mexico in the coming weeks.  She received her
M.S. in 1996 and her B.A. in Literature from Duke University in 1990.
Her current research interests are in operating systems, networking
and high-performance computing.  She has worked in the past with
industry, the national science laboratories, and academia in the areas
of Internet performance, digital television, and multi-agent security.