Hardware Reference
In-Depth Information
Fig. 3.1
An example of
Pareto-dominance in a
two-objective minimization
problem. The circles represent
non-dominated points, while
squares are dominated points
0.60
0.50
0.40
0.30
0.2
0.10
0.20
0.30
0.40
Minimize x
In problems with discrete categorical configuration parameters, the types of the
input variables are discrete ranges and might also be unordered collections, meaning
that optimization methods which assume an order relation cannot be used profitably.
For example, the number of processors on the platform is a discrete ordered variable
since it is a natural number. Instead, the type of
branch predictor
to be used (e.g.,
static, two-level adaptive, etc.) is a categorical variable since an ordering between
the different instances cannot be defined.
The evaluation of the objective values for the designs selected for exploration
is usually performed through a simulator during the optimization phase. Simulators
accuracy depends on their level of abstraction which is inversely proportional to their
computational complexity. A manual exploration procedure might include additional
delays to the already long simulation time.
Two examples of automatic optimization frameworks will be considered in this
chapter. The first framework is the open source Design Space Exploration (DSE)
Multicube Explorer. It has been initially conceived for the kind of problems dis-
cussed above. Throughout this chapter, a description of its optimization algorithms
introduced within the framework will be provided.
The second tool is modeFRONTIER, a commercial software which has been
widely used worldwide for more than ten years in different domains like aerospace,
appliances, pharmaceutics, civil engineering, manufacturing, marine multi-body de-
sign, crash, structural, vibro-acoustics and turbo-machinery. All these domains define
multi-objective optimization problems, but in continuous or mixed (continuous and
discrete) domains, not in complete discrete and possibly categorical domains like the
SoC design problems. Due to this reason, an initial re-target process to add support
for categorical variables to modeFRONTIER has been carried out and the actual
release of the software contains this work as well as the algorithms developed within
the project.
The automatic design space exploration performed by one of these two tools is
governed by an optimization algorithm. The algorithm is responsible for choosing the
new configurations which have to be simulated and for analyzing the results obtained.
The optimization phase can be preceded by a Design Of Experiments (DOE) study
and it can be followed by some Post Processing analysis.
Search WWH ::
Custom Search