MIME-Version: 1.0
Server: CERN/3.0
Date: Sunday, 24-Nov-96 22:57:13 GMT
Content-Type: text/html
Content-Length: 6359
Last-Modified: Thursday, 21-Dec-95 19:08:30 GMT
Masters of Engineering Projects
This year all our projects will address issues associated with
the development of an environment for designing, specifying,
implementing and testing Adaptive Mesh Refinement (AMR)
methods for tackling very complex problems (ex. 3D spiraling
coalescence of two black holes) from computational sciences.
The participants in the development of this environment will have
the opportunity to work on very specific projects that can be developed
individually, but at the same time will require some degree of interaction
in the design, validation and testing phases. Thus, in addition to the
opportunity of acquiring specific skills and knowledge by working
on a very narrow and interesting problem the opportunity to see and
understand the procedure of designing and implementing a complex
software for parallel and distributed platforms will be given.
Projects :
Intelligent Graphical User Interface for AMR (GUI-AMR)
This project requires the design and implement an intelligent
frond-end for specifying 2D geometries, operators and various
parameters required to solve numerically, time-dependent PDE problems.
This project consist of two components:
A. Problem specification
B. Visualization
Existing Problem Solving Environment (PSE) PELLPACK will be extended.
This project is recommended for students with interest on GUIs,
PSEs, I/O, and in general front-ends for high-performance computing
environments.
Parallel Runtime Support System for AMR (PRTS-AMR)
This project requires the design and implementation
of communication and threaded modules requirted for
the efficient implementation of data-movement and
control primitives for parallel AMR methods.
Existing communication software like MPI, AMs and threaded
packages like Qt and PORTS0 will be utilized.
This project is recommended for students with interests
both on systems, parallel I/O, parallel compilers and computational
sciences. An interaction with Bernoulli and Split-C projects
is expected.
Dynamic Load Balancing Algorithms for AMR (DLB-AMR)
This project requires the development, implementation and evaluation of
of dynamic load balancing algorithms for AMR: (i) direct like generalized
spectral bisection and its derivatives and space-filling curves
and (ii) incremental. The objective is to develop (i) a DLB-AMR module
for PRTS-AMR and (ii) acquire useful knowledge in solving an instance
of a very difficult combinatorial optimization problem: dynamic load
balancing of adaptive computations on a network of time/memory-sharing
heterogeneous workstations and MPPs. Existing algorithms and software
will be extended and new ones will be build. This project is recommended
for students with interests on the solution of practical optimization
problems related to parallel and distributed computing.
Parallel PDE solvers for time-dependent problems using AMR methods
This project requires the implementation of a library of routines for the
discretization and solution of 2D/3D wave equation on MPPs and SMPs.
Both non-threaded and multithreaded paradigms will be used and evaluated.
In addition accuracy, and stability of multithreaded computations will be
analyzed. High-performance languages like HPF and low level like
Fortran/C plus message passing will be used. This project is recommented
for students with interests in parallel numerical computing, and
scalability analysis. An interaction with PRTS-AMR project is expected.
Parallel structured grid generation for 2D/3D complex geometries
This project requires the development of parallel algorithms and
implementation of a library of parallel grid generation modules using :
- Algebraic Methods
- Elliptic Methods
- Moving Algebraic Methods
High-performance languages like HPF and low level like Fortan/C plus
message passing will be used. This project is recommented for students
with interests in parallel numerical computing. Existing state-of-the-art
scalar algorithms and code will be used. An interaction with WLIB-AMR is
expected.
Scientific computing benchmarks for multithreading on parallel and
distributed platforms
Study the impact of restructuring CG scalar algorithms and codes in terms of CG's performance on MPPs
This project requires the development of a Conjugate Gradient algorithm
suitable for multiprocessors and includes the implementation of CG using
Typhoon parallel compiler. Issues related to expensive dot product
operations will be studied:
- How much the algorithm restructuring can help in improving
performance compared to well known CG algorithm?
Is it worth the effort and possible increase in space complexity?
- How much the code restructuring can help in improving
performance compared to straight forward cg code?
Is it worth in doing the extra step in exploring functional
parallelism on top of the data-parallelism?
- How much will cost to use multithreading in (i)
implementing functional parallelism (ii) masking some of the
global address space overheads? Is the overhead worth the
increase in software complexity? What about error and stability?
For more information contact: