ON THE ISSUES OF APPLICATION PERFORMANCE IN PARALLEL AND 
DISTRIBUTED SYSTEMS FROM PC CLUSTERS TO DESKTOP GRIDS

Prof. Thomas M. Stricker
Swiss Federal Institute of Technology, ETH Zuerich

ABSTRACT:

Designing compute platforms with good scalability and optimal price
performance for a given application remains a significant challenge as
we are still lacking a systematic understanding of the architectural
issues around precise resource usage by many applications on those
platforms. To address this problem, we derived a performance analysis
framework for filtering and abstracting the results of performance
samplers or performance counters into a higher level picture that
reflects the overall application performance.

We demonstrate the viability of our approach as we look into
distributed TPC-D running on clusters of commodity PCs. We isolate and
resolve a few crucial performance issues. In particular we can give a
workload characterization for each query in terms of precise resource
usage, quantify its scalability and investigate e.g. the impact of the
networking speed.  Considering a different application domain we 
investigate the performance of common molecular dynamics simulations and
show how to measure the performance on existing platforms, model it
and extrapolate the resulting figures to some novel platforms for
widely distributed computing on desktop grids.


SPEAKERS BIO: 

After growing up in Switzerland, going to school there
and getting slightly bored there Thomas Stricker joined the staff of
IBM T.J.Watson research lab 1988. Working on CAD tools for EMI, he
became gradually interested in computer science research.  He moved
from New York to Pittsburgh to attend the Ph.D. program of Carnegie
Mellon University, where he contributed to iWarp parallel computer
architecture project, the Fortran Fx compiler project and occasionally
consulted for an NSF grand challenge on the best architecture to use.
After defending a doctoral thesis on "direct deposit - where message
passing meets shared memory" - he moved back the ETH Zuerich where he
is currently an assistant professor. The research focus of his group
there is on systems architectures from clusters of commodity PCs to
widely distributed computing on Internet.