narayan@atanasoff.cs.iastate.edu (Pankaj Narayan) (02/05/90)
I have a few questions and need a little input on what will
(I hope) turn out to be my MS Thesis.
I propose to study parallelism with an emphasis on how it can
be benificial to graphically-intensive tasks. To begin with, there
needs to be some kind of a breakdown of the various issues
that arise once you try to break up a task so as to do it in parallel;
for example, HOW to break it? do we worry about dependencies (of
one part of the break-up on another) ? etc....
Once I have explored in sufficient detail (which is where the problem
is; where do I START ??) all these branches of the tree, I want to
then take up a problem that is of some practical value and fairly
computational and graphically intensive (e.g. something
in Computational Flow Dynamics, like tracing a particle in motion
in a flow field) and actually implement the parallel version of
it (in X, to take advantage of the network-independent features
it provides) to show the speedup observed.
Also, I wish to show that despite a speedup in the computation,
the bus that takes the results to the screen (to display in a
graphical format) acts like a bottleneck for the "how fast
can I do this?" question.
So we propose a system whereby each of the processors (of the
parallel machine) be assigned one graphics screen, and that the
results computed by that processor be displayed on that screen. The
screens can then be placed side-by-side to give the final overall
result. (This would force us to distribute tasks in the same manner
as we wish to view the results, which translates to breaking
up a data set into well-defined chunks and assigning each chunk to
a separate processor; also forces us to pick such a problem
to parallelize that has no inter-data-field dependencies, so that
the break-up may be clean)
I need some inputs on a few things.......
- has any of you ever seen work thats been done to throw light
in this area of study of the different ways of parallelizing
an application ?
- does anyone have any applications in mind that have been
written for a computational and graphical intensive task
(preferably using the X window protocol) but targeted
for a sequential architecture, but you think would be a natural
candidate for parallelizing ?
- any comments on any other issues you feel are worth exploring
are more than welcome, since I am just starting out on my thesis
and so am groping around trying to define its scope...
Please post or mail, whichever is more convenient.
Pankaj Narayan narayan@atanasoff.cs.iastate.edu
246 N Hyland Apt. 306 Ames, IA 50010
515 292 5535kyriazis@herodotus.rdrc.rpi.edu (George Kyriazis) (02/06/90)
In article <7883@hubcap.clemson.edu> narayan@atanasoff.cs.iastate.edu (Pankaj Narayan) writes: > > So we propose a system whereby each of the processors (of the > parallel machine) be assigned one graphics screen, and that the > results computed by that processor be displayed on that screen. The > screens can then be placed side-by-side to give the final overall > result. (This would force us to distribute tasks in the same manner Take a look at the SIGGraph 1989 proceedings, the volume on 'Hardware Advances in Computer Graphics'. I agree with you that graphics can ne a bottleneck (especially in the future), but your approach will probably have granularity problems. Interleaved frame buffers usually perform better. I think this book will answer many of your questions. Some other volumes of the same proceedings can also been enlighting (they describe different algorithms or implementations of common problems). > Pankaj Narayan narayan@atanasoff.cs.iastate.edu > 246 N Hyland Apt. 306 Ames, IA 50010 > 515 292 5535 George Kyriazis kyriazis@turing.cs.rpi.edu kyriazis@rdrc.rpi.edu ------------------------------