Revised Time: Ripeanu/M.S. Presentation/Nov. 15

[Margaret Jaffey]

The time of Matei Ripeanu's Master's Presentation on Wednesday, 
November 15th, has been changed.  It is now scheduled to begin at 
9:00 a.m. (instead of 1:00 p.m.).  All other details remain the same. 
A copy of the announcement is included below.


Date:  Wednesday, November 15, 2000

Time:  9:00 a.m. (revised time)

Place:  Ryerson 275

M.S. Candidate:  Radu Matei Ripeanu

M.S. Paper Title:  Issues of Running Large Scientific Applications in a
		Distributed Environment

Abstract:

The CACTUS software can be considered representative for a whole class
of scientific applications: tightly coupled, regular space
decomposition, huge memory/processor time requirements.  This paper
tries to address issues related to the execution of such applications
beyond the traditionally used resource environments: single processor,
shared memory and message passing multiprocessors or network of
workstations.  We study the requirements of running Cactus on two new
resource environments: a pool of geographically distributed
supercomputers and a ‘mega?computing’ environment.  We first derive
application performance models for two traditional architectures
(Silicon Graphics Origin 2000 and IBM SP) and validate them against
experimental data.  Then we draw up performance models for the two
environments introduced and present evidence that our models are valid.
Analyzing our models and experiments we discovered (at least) two areas
that should be investigated further: load balancing in highly
heterogeneous environments and the scaling of the transport protocol.

One characteristic of these environments is heterogeneity; therefore it
is very likely that processing units will have different processing
powers.  Due to the tightly coupled structure of the application this
means that the whole application will advance at the speed of the
slowest processor.  To offset this we implemented a dynamic partitioning
scheme of the grid space that allows load balancing.  We propose a
metric to evaluate partitioning algorithms and present experimental
results that show that even a naïve partitioning heuristic produce large
benefits when running Cactus.

TCP has been widely used as transport protocol even in high performance
computing communities.  We analyze the benefits and limitations of using
TCP in environments in which thousands of simultaneous flows are likely
to compete for bandwidth.

Mr. Ripeanu's Advisor:  Prof. Ian Foster

A draft copy of Mr. Ripeanu's M.S. paper is now available for review 
in Ry 161A.

Margaret
-- 
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey			margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 161A)		(773) 702-6011
The University of Chicago		http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=