Have project idea, seeking partner(s)
Mark Frederick Augustine Hoemmen
Am seeking project partners relating to a particular form of domain
decomposition for finite difference methods, called the ``ghost cell
expansion'' (GCE) method. In finite difference schemes where the region
of interest is partitioned into domains, wave propagation across the
boundary requires communication, so that the values of ``ghost cells''
(grid points or cells on the other side of the boundary) can be used in
the finite difference stencil.
The authors of a paper I've read (see BibTeX entry at the bottom of this
e-mail) propose increasing the width of the boundary region beyond what
the finite difference stencil needs, so that communication can be
delayed for a few iterations and then combined into larger messages.
This reduces overall communication and is favorable for networks that
fit the LogGP model (rewarding large messages), but the benefits differ
widely for different networks. Furthermore, the technique can affect
the accuracy of solution (which is something I'm looking at for a Math
228B project). So there is a potential for tuning.
I'd like to find at least one partner who uses finite difference methods
(or at least some numerical PDE solution technique) in his or her
application area.
The BiBTeX entry for the paper which inspired this idea can be found at
the end of this e-mail; look in http://portal.acm.org/ for the full
text.
mfh
@InProceedings{ding2001ghost,
author = {Chris Ding and Yun He},
title = {A ghost cell expansion method for reducing communications in
solving PDE problems},
booktitle = "Proceedings of the 2001 {ACM/IEEE} conference on
Supercomputing",
month = nov,
year = {2001},
publisher = {ACM Press},
address = {New York, NY, USA},
isbn = {1-58113-293-X}
}