The above analysis indicates that, for fixed p and k, the communication complexity is independent of the problem size. Hence, as n increases, we expect the local computation to dominate.
The histogramming algorithm has been implemented on a CM-5 with
p=16, 32, 64, and 128 processors, and the algorithm's
performance is plotted in Figure 3 for 256 grey
levels for images ranging from
to
pixels in size, and expanded in
Figures 12 - 14 for
to
images on 16, 32, and 64
processors. Corresponding performance graphs are given for the IBM
SP-1 in Figure 18 and for the IBM SP-2 in
Figure 20. Plots indicate quadratic performance
as a function of n for fixed p, and scalability in terms of p.
Hence, our theoretical analysis is supported.
Figure 3: Histogramming and Connected Components Scalability on the CM-5
Please refer to the plot in Figure 3 for an
illustration of the scalability of the histogramming algorithm's
performance. Since computation dominates for large n, the algorithm
runs as O
. We have plotted
vs. time for
four configurations of the CM-5. The resulting plot shows the linear
relationship between time and image size for each fixed p. Also,
when the number of processors double, the running time approximately
halves.