Design Space Exploration of Parallel Architectures - Multi-objective Design Space Exploration of Multiprocessor SoC Architectures - page 176

14

8

12

10

6

8

4

6

4

2

2

0 0

32 48 64 80 96 112 128 32 48 64 80 96 112 128

rob_depth rob_depth

a b

Random Uniform Distribution Uniform Latin Hypercube Distribution

Fig. 8.2 The marginal distribution for the parameter rob_depth with a a “Uniform Random” and b
“Uniform Latin Hypercube” Design of Experiments with 60 designs. It can be appreciated that the
distribution of samples for each value is better in b , showing that the second approach has provided
a better sampling of this discrete design space

way, that the time and resources needed to make the experiments is reduced as much
as possible.
In these experiments, a Uniform Latin-Hypercube algorithm [ 4 ] is selected to
generate 60 initial designs. The selection of the DoE algorithm and the number
of designs are driven by the fact that only 8,134 designs can be evaluated and as
described in Chap. 3, this algorithm is preferred over the widely-used Uniform
Random algorithm, particularly due to the better mapping of the marginal distribution
functions with a small number of samples, as clearly shown in Fig. 8.2 .

8.2.3.3 Optimization

The algorithm MFGA (Magnifying Front Genetic Algorithm) [ 7 ], described in Chap.
3, has been selected by considering the characteristics of the optimization problem.
This algorithm has been developed in the MULTICUBE project to handle categorical
discrete optimization problems. A non-generational operation mode (steady state)
makes it well suited for problems involving long simulation time. This algorithm
has only two configuration parameters, which are set to default suggested values:
crossover probability to 0 . 9 and mutation probability to 0 . 15.
The optimization is performed by using this algorithm till the allowed number
of evaluations is reached without any manual intervention. Figure 8.3 shows the
design space evaluation, highlighting the Pareto front as obtained at the end of the
evolutionary process 1 . The values of the metrics (normalized due to confidentiality
reasons) are plotted for each evaluated design with a bubble which size is directly

1 Even if unusual, a bubble graph with the third dimension as the bubble size was preferred over a
typical 3D graph since the large number of points produces a difficult to understand 3D cloud-like
graph