CISE Technical Lecture Series
CISE -TLS: August - December 2003
CISE Lecture I - Dr. Nayda G. Santiago Santiago August 28, 2002
Evaluating Performance Information for Mapping Algorithms to Advanced
Architectures
The development of efficient
code for scientific and engineering applications on advanced computing systems
is not a trivial task. To accomplish this task, a code developer has to be concerned
not only about algorithmic correctness and robustness, but also about performance
and implementation details. These additional factors impose a burden on the typical
scientific computing expert, preventing the user from effectively leveraging
the computational resources available to the application. Two major factors can
be identified among those making this task particularly difficult. First, the
complex interactions between the target platform and the application software
tend to hide information about the existing relations between different entities
in the system. Second, the high dimensionality of the performance data conceals
interesting patterns in the observations which could lead to insights into the
system behavior. While a multiplicity of tools have been developed to solve these
problems, many obstacles still exist when characterizing the relations among
high-level factors and low-level performance information. These problems not
only make difficult the task of efficient coding, but also prevent the development
of automated performance analysis tools to assist application programmers to
tune their code. In this presentation we will present a new methodology
for obtaining information about the relations emerging when compute-intensive
applications are mapped onto advanced architectures. The proposed methodology
incorporates knowledge and techniques from multiple areas that include statistics,
operational research, pattern recognition, data mining, and performance evaluation
to enable the extraction of performance information during the mapping process.
The methodology is composed of four steps: problem analysis, design of experiments,
data collection, and data analysis. In the first two steps, analyses of the application
itself are completed to determine the appropriate design of experiments for establishing
relations between changes in high-level abstractions and performance outcomes.
Feature subset selection is proposed for identifying important system metrics.
An evaluation of different statistical analysis alternatives was carried out
to characterize the types of data obtained in performance studies. The information
obtained from this methodology can be converted into appropriate suggestions,
observations, and guidelines for the scientific computing expert to tune applications
to a particular computing system.
November 6, 2003
Elastically Replicated Information Services (ERIS)
Elastically Replicated Information Services (ERIS) use adaptive replication
algorithms in order to sustain a desired level of data availability even in
the presence of online changes to the topology of a distributed storage system.
In this talk we first argue for the need to achieve a dynamic balance among
data replication and data migration in order to sustain a desired level of
availability while maximizing storage utilization. The talk will also illustrate
the inherent trade off between utilization and availability in distributed
storage systems and introduces a simple mathematical model of a distributed
storage cluster (DSC). Preliminary results from simulations of different elastic
replication algorithms using this DSC model suggest that ERIS are desirable
even in small scale DSC's (order of ten storage nodes). Our initial empirical
results provide an early indication of a practical need for elastic replication
algorithms.
CISE Lecture II -Dr. Hugh B. Nicholas Jr.
November 20, 2003
Predicting the Determinants of Enzyme Specificity: Combining Biology, Mathematics, and Computer Science in Molecular Biological Research.Present day biological organisms possess many homologous gene families (i.e., gene families that share a common evolutionary ancestor) that encode proteins that carry out similar but distinct biochemistry and participate in different physiological processes. These different proteins frequently function in the same cellular environment. Thus, the protein (and gene) sequence must encode the information that allows molecules to participate in the processes and pathways to which they are specific and to also avoid participating in "incorrect" processes and pathways. I will present analyses that identify the residues in biological macromolecules that are most likely to confer the specificity of interaction described above. The talk will first describe the analyses in terms the essential biochemical and biological properties of the molecular systems being studied and then discuss encapsulating this biological knowledge into formal mathematical models of the biological system. Finally, the mathematical model into a computer algorithm, and program, that calculates how closely specific biological sequence data conforms to this model. The talk will describe applying the analysis to macromolecular families and will look at how well these computational predictions match experimental results addressing the same biological questions. I will conclude by outlining some open mathematical and computational questions related to the analysis.
SPECIAL EDITION
CISE Technical Lecture Series - SPECIAL EDITION September 4, 2003 Dr. John RodriguezSilicon Technology Development - Texas Instruments
Reliability Aspects of Ferroelectric Memories
A low cost, low power, non-volatile memory technology is required for many
applications. One promising candidate utilizes the ferroelectric material PZT.
Ferroelectric films can be polarized in either of two stable states and the
polarization remains even when the writing voltage is removed, making ferroelectric
memories non-volatile. In addition, ferroelectric memories can operate at low
voltages, making them attractive candidates for integration with state of the
art CMOS processes. In this talk, we will briefly review the basics relevant
to ferroelectric memories, including electrical characterization. The two primary
reliability concerns for FRAM are fatigue due to bipolar cycling and data retention.
We will describe these effects and highlight progress achieved in improving
the reliability of these films.