Transcript Parallel and Distributed Computing Research at the Computing

Parallel and Distributed
Computing Research at the
Computing Research Institute
Ananth Grama
Computing Research Institute and
Department of Computer Sciences
Purdue University
http://www.cs.purdue.edu/people/ayg
Areas of Research
• High Performance Computing Applications
• Large-Scale Data Handling, Compression,
and Data Mining
• System Support for Parallel and Distributed
Computing
• Parallel and Distributed Algorithms
High Performance Computing
Applications
• Fast Multipole Methods
– Particle Dynamics (Molecular
Dynamics, Materials Simulations)
– Fast Solvers and Preconditioners for
Integral Equation Formulations
– Error Control
– Preconditioning Sparse Linear
Systems
• Discrete Optimization
• Visualization
System Support for Parallel
and Distributed Computing:
• MOBY: A Wireless Peer- to- peer Network
• Scalable Resource Location in Service Networks
• Scheduling in Clustered Environments
Large-Scale Data Handling,
Compression, and Mining
• Bounded Distortion Compression of Particle Data
• Highly Asymmetric Compression of Multimedia Data
• Data Classification and Clustering Using Semi-Discrete
Matrix Decompositions.
Parallel and Distributed
Algorithms
• Scalable Load Balancing Techniques
• Parallel Programming Paradigms
• Metrics and Analysis Frameworks
(Isoefficiency, Architecture Abstractions
for Portability)
Computational Elements of
Robust Civil Infrastructure
• Civil infrastructure represents the single largest
investment in the United States, valued at over $20
trillion.
• While these systems are in a constant state of
renewal, they are often required to withstand extreme
loads caused by natural disasters or human
intervention.
• High-rise structures, long-span bridges, dams, and
pipelines are particularly vulnerable.
• The serviceability and safety of these structures can
be vastly improved if damage can be detected and
controlled in real-time.
Computational Elements of
Robust Civil Infrastructure
• With the availability of reliable inexpensive sensors,
large-scale actuation devices, and computing and
communication elements, the technology for active
control of large structures exists, in principle.
• The goal of this ambitious project is to:
– Enable effective design and economical
construction of highly robust smart structures.
– Enhance robustness of existing structures by
suitably retrofitting them.
– Predict and mitigate impact of catastrophic events,
– Provide support for area-wide disaster
management plans.
State-of-the-art in Controlled
Structures
Building Blocks of Smart Structures
Magnetorheostatic dampers can change
their load bearing characteristics from fully
solid to fully damping in milliseconds
when exposed to magnetic fields.
Sensing/Computation/Communication
elements - designed by part of our
research team at Dartmouth. These units
cost under $200 and are the size of a deck
of cards. This is a rapidly evolving field and
efforts are on to develop the next
generation of devices here at Purdue.
Control Timelines
Control Strategy
Outstanding Challenges
• Building reliable inexpensive
sensing/computation/communication/actuation
(SCCA) units.
• Building a reliable network of SCCA units.
• Structural modeling and model reduction.
• Execution of the distributed control algorithm
with tight real-time constraints.
• Supporting an area-wide disaster
management information network.
Computational Aspects of Multiscale Modeling of NEMS
•
•
•
•
Efficient Numerical Algorithms
Parallel and Distributed Computing
Software and Libraries
Interfaces to Experimental Data Acquisition
and Design Components
• Interfaces to Application Servers
The overall goal is to develop a comprehensive simulation
environment built upon novel algorithms and parallelism for multiscale modeling of NEMS.
Technical Objectives
Technical Challenges
•
•
•
•
•
Diversity of phenomena - multiphysics
Variance in spatial scales - nm to cm
Variance in temporal scales - fs to s
Variety of modeling phenomena
Self consistency between scales and
phenomena
Technical Challenges
Computational and
Mathematical Challenges
• Novel problems in linear algebra
• Special functions and approximations
• Self consistency between scales and
phenomena
• Highly dynamic geometries and interfaces
• Extremely large number of degrees of
freedom
• Need for scalable parallelism
Collaborations
• Structures: Mete Sozen, Robert Frosch
• NEMS, Networks and Control: Mark
Lundstrom, Supriyo Datta, Kent Fuchs, Jim
Krogmeier, Mark Bell, Ness Shroff, Rudi
Eigenman
• Laser Ablation: Jayathi Murthy, Xianfan Xu
• Algorithms and Software: Ahmed Sameh,
Chris Hoffmann, Sonia Fahmy, Zhiyuan Li