Title: CS6035 Parallel/Distributed Processing II:
1CS6035 Parallel/Distributed Processing II
An Efficient Static Assignment Parallelization
Scheme for Algebraic Fractals
By Chris MacPhee Supervisor Dr. Bhavsar
2Outline
- Introduction
- Computational Characteristics
- Serial Program
- Parallelization
- Experimental Results
- IBM SP
- SGI Onyx
- Conclusion
3Introduction
What are fractals?
- Possess non-Euclidian geometry (formless)
- Self-similar (same type of structure at all
scales) - Clouds are not spheres, mountains are not cones,
coastlines are not circles, and bark is not
smooth, nor does lightning travel in a straight
line. - - Benoit Mandelbrot, 1983
4Introduction
Examples of fractals images
From Fractal Gallery http//projekt.pinknet.cz/f
ractal/
5Computational Characteristics
The Mandelbrot set
- z ? z2 c , where z, c ? ?
- z0 is a constant
- c varies
- z is iterated until either
- z diverges beyond a preset limit
- the maximum number of iterations
- is reached
6Computational Characteristics
The Mandelbrot set (cont)
z ? z2 c
7Computational Characteristics
The generalized function
z ? z2 c , where z, c ? ? z ? z? c ,
where z, c ? ? ? ? ?
8Computational Characteristics
The generalized function (cont)
z ? z10 c
9Computational Characteristics
The generalized function (cont)
z ? z100000 c
10Computational Characteristics
Distribution of iterations
11Parallelization
Two architectures
- Shared memory programming
- Run on SMP machines (e.g. Sun SGI)
- Uses OpenMP
- Message passing programming
- Run on distributed memory machines
- (e.g. Compaq IBM)
- Uses Message Passing Interface (MPI)
12Parallelization
Three work assignments
Static work assignment Dynamic work
assignment New static work assignment
13Parallelization
Static work assignment
Divide column groups evenly between processors
Master
257-512
1-256
513-768
769-1024
14Parallelization
Dynamic work assignment
Farm work to the slaves in work sizes of 64
columns
Master (in queue 513-1024)
449-512
321-384
385-448
257-320
15Parallelization
New static work assignment
Divide workload evenly over processors
Master
370-513
1-339
514-657
658-1024
16Experimental Results
Two machines
- Symphony (University of New Brunswick)
- IBM SP
- 16 ? 375 MHz processors
- 4 GB of RAM
- Distributed memory architecture
- Herzberg (Memorial University of Newfoundland)
- SGI Onyx
- 28 ? 400 MHz processors
- 14 GB of RAM
- Shared memory architecture
17Experimental Results
IBM SP Timings
Computing time for each slave processor for ? 2
18Experimental Results
IBM SP Timings
Computing time for each slave processor for ?
10
19Experimental Results
IBM SP Timings
Computing time for each slave processor for ?
100000
20Experimental Results
SGI Onyx Timings
Computing time for each slave processor for ? 2
21Experimental Results
SGI Onyx Timings
Computing time for each slave processor for ?
10
22Experimental Results
SGI Onyx Timings
Computing time for each slave processor for ?
100000
23Summary
Summary
- The computational characteristics of fractal
images have been analyzed. - A static assignment method for efficient parallel
processing has been developed. - The static assignment method becomes more
efficient as ? increases.
24References
1 H. O. Peitgen and P. Richter, The Beauty of
Fractals, Springer-Verlag, Berlin,
1996. Â 2 U. G. Gujar and V. C. Bhavsar,
"Fractals from z z a c in the Complex
z- plane", Comp. and Graph., 16(1), pp. 45-49,
1992. Â 3 S. V. Dhurandhar, V. C. Bhavsar, and
U. G. Gujar, "Analysis of z-plane fractal images
from z z a c for a lt 0", Comp. and Graph.,
17(1), pp. 89-94, 1993. Â 4 V. C. Bhavsar, U.
G. Gujar, N. Vangala, "Vectorization of
generation of fractals from z z a c on IBM
3090 / 180VF", Comp. and Graph., 17(2), pp.
169-174, 1993. Â 5 E. Aubanel, "Parallel
Programming with Generalized Fractals," Faculty
of Computer Science, University of New
Brunswick, February 2002, http//www.cs.unb.ca/p
rofs/aubanel/aubanel_fractals.html. Â 6 B.
Wilkinson and M. Allen, Parallel Programming,
Prentice Hall, Upper Saddle River, 1999.