April 29, 9:00 - 10:00 am
Introduction to High Performance Computing
Brian Dobbins
In recent years, clusters of commodity PCs
have gone from being a small experiment at NASA's Goddard Space Flight
Center to a dominant paradigm of research computing in universities and
industries across the globe. This talk will cover the origin of Beowulf
clusters, their benefits and limitations, where they're commonly used,
and what they offer a computational researcher.
April 29, 10:00 - 11:00am
Cluster Hardware
Brian Dobbins
Description coming soon
April 29, 11:00-12:00 pm
Cluster Software
Brian Dobbins
Description coming soon
April 29, 1:00-2:30 pm
Parallel Programming with MPI
Angelo Rossi, IBM
Description coming soon
April 29, 2:30 - 3:30 pm
Parallel Programming with LINDA
Nick Carriero, Yale Computer Science
Description coming soon
April 29, 3:30- 4:00 pm
Yale ITS Research Clusters
Paul Gluhosky, Yale ITS
Description coming soon
Abstract coming soon
How do cortical neurons work together
to encode information? Using statistical pattern
recognition approaches, we compared the informational content of
individual neurons with their contribution to the information encoded
by an ensemble of neurons. We demonstrate that the information content
of a single neuron is not related to its contribution to an ensemble of
neurons; that is, the neurons which best predict a stimulus or behavior
independently are not necessarily those whose addition or removal most
changes the ensemble's predictive ability. Furthermore, we find
that our neuronal ensembles are highly redundant, and maintain
performance as they are degraded. We related the
redundancy in our neuronal ensembles to correlated noise.
Finally, we find that when the complexity of the predictive information
becomes high relative to the number of neurons, we observe that small
ensembles of neurons may be synergistic. Taken together, these
results suggest that neural ensembles exhibit distributed coding.
This talk describes the Linux Blade Cluster in
use by the High Energy Physics Group at Yale. A brief description of
the cluster hardware and software is followed by 3 researchers
describing how the cluster is being used in their research. The
presenters provide information on three research projects currently
using the cluster. Alpha Magnetic Spectrometer (AMS), Quasar Equatorial
Survey Team (QUEST), and CDF (Collider Dectector Facility).
Simulations of biologically realistic neurons are generally
run thousands of times with only a few parameter
changes. Since a single run almost always takes longer than 0.1
seconds the problem of utilizing multiple machine resources is entirely
administrative. For these embarassingly parallel problems NEURON
offers a bulletin board on which todo tasks are posted. As long
as the bulletin board is non-empty, no host waits for other hosts'
results, but constantly takes a task and puts the result back onto the
bulletin board.
I will discuss some of the unique
processing requirements for a humanoid robot that we have been
developing at Yale for the past two years and some of the approaches
that we have tried to solve these problems. Our robot generates more
than 200 megabytes of data per second, must control 30 actuators, and
utilizes a 20 processing node computational system composed of
off-the-shelf components connected with a rich network of both
point-to-point and broadcast network connections.
Quantum and statistical mechanics are used to
model reactions in solution and protein-ligand binding. The
computational demands have been met by use of large PC clusters running
Linux. Details will be provided on our cluster, the nature of the
calculations, and the significance of the results.
No abstract available
Biomedical domains are increasingly
relying on globally distributed information repositories due to the
significant surge in the amount of data being generated. The emerging
Grid technologies seem to provide a viable solution to the on-demand
integration/extraction of/from massive biomedical data sets due to the
advances in network bandwidth, storage capacity, and internet
technology, etc. However, the current Grid middleware developments,
such as the Globus, Legion, and the UNICORE, do not well provide the
ability to be on-demand, interactive, as well as to securely manage and
share data to the metadata level. In this talk, we will discuss the
efforts in the past, present, and future of the Grid research community
to support potential biomedical applications.
After a short introduction, we will demonstrate STAR-P
April 30, 1:30 - 2:30 pm (Dunham 220)
Title Not Yet Determined
Gregory Warnes, Pfizer Inc.
Abstract not yet available
Using high-resolution SPH numerical simulations, we investigate the effects of gas on the inspiral and merger of a massive black hole binary. This study is motivated by both observational and theoretical work that indicate the presence of l arge amounts of gas in the central regions of merging galaxies. N-body simulations have shown that the coalescence of a massive black hole binary eventually sta lls in a stellar background. However, our simulations suggest that the massive b lack hole binary will finally merge if it is embedded in a gaseous background. Our work thus supports scenarios of massive black hole evolution and growth wher e hierarchical merging plays an important role. The final coalescence of the black holes leads to gravitational radiation emission that would be detectable up t o high redshift by LISA.
Many important and interesting phenomena can be modeled by convolution. For example, atmospheric distortion of the image of a celestial object can be modeled by the convolution an exact but unknown image function with a "blurring" function. We are often interested in deconvolution--the inverse problem associated with undoing the convolution. In the astronomy example, we would like to recover the exact image of the celestial object from the observed distorted image. Deconvolution problems, like inverse problems in general, are theoretically and computationally difficult to solve. Solutions are extremely sensitive to perturbations in the observed data (e.g., noise, numerical representation errors), requiring some form of regularization. Regularization replaces the deconvolution problem with a related problem that is less sensitive to perturbation. Computational challenges arise from the often enormous size of the problems resulting after discretization.
Traditional approaches to solving deconvolution problems involve Tikhonov regularization and basic iterative refinement methods (Richardson iteration). More recently, various Krylov subspace iterative methods have found application to the regularization of deconvolution problems. The Krylov subspace methods typically exhibit much faster convergence and computational efficiency than the earlier methods. We will discuss Krylov subspace methods for deconvolution using Tikhonov and other forms of regularization. Implementations including methods using parallel computation, will be discussed. Ideas will be illustrated with practical examples from astronomy and confocal microscopy.
"Real world" performance problems tend to arise in a context
in which one or more or the following apply:
1) The code in question is an evolving specification
of a research group's solution technique.
2) The code in question is, effectively, a black box.
3) There are additional processing steps which
establish both compatibility requirements and a performance floor,
beyond which performance improvement is overkill.
Such problems are poorly served by the pedal-to-the-metal mentality
that is an occupational hazard of HPC. We discuss several case studies
illustrating these points and describe a variety of techniques that
were used to achieve adequate levels of performance.
No abstract available
