Experimental Parallel Algorithmics
Experimental Data Sets:
Dynamic Data Redistribution
Institute for Advanced Computer Studies
University of Maryland, College Park
Variables |
|
|---|---|
| n | total number of elements |
| p | number of processors |
| i | the processor number from ( 0 <= i <= p-1) |
| m | the maximum number of elements initially on a processor |
Balanced
Linear
Normal
Exponential
All-on-one
Each processor initially holds n / p elements and hence m = n/p.
Each processor initially holds i × [ 2n / (p(p-1))]
elements and hence m = 2n/p.
Elements are distributed in a Gaussian curve and hence m is
approximately 2.4 n/p for p >= 8. A code snippet appears
here.
Note: we sample a mean zero, s.d. one, Gaussian curve at the center of
p intervals equally spaced along [-3,3]. The sample values are
normalized to sum to n by multiplying each by n / (sum of
the p samples).
Processor i contains n / 2^{i+1} elements, for
i < p-1, and processor p-1 contains
n / 2^{p-1} elements and hence m = n/2.
An arbitrary processor contains all n elements and hence m=n.
Return to the Experimental Parallel Algorithmics Experimental Data Sets page.
dbader@umiacs.umd.edu
Balanced
Linear
Normal
Exponential
All-on-one
