Design Space Exploration of Parallel Architectures - Multi-objective Design Space Exploration of Multiprocessor SoC Architectures - page 174

Table 8.2 The system metrics generated by the simulator
Category Metric Description

Performance total_cycle Total cycle number

total_inst Total instruction number

ipc Instruction per cycle

Power dissipation total_energy Total energy consumed

power_dissipation Average power dissipation

peak_power_dissipation Peak power dissipation

Area area Area occupied

8.2.3 Design Space Exploration

The aim of the experiments presented in this section is to determine the set of designs
that minimize the total number of cycles, the power dissipation and the area occupa-
tion required to run the selected gzip application with the SP2 simulator. Since it is a
multi-objective problem, and since the three objectives can potentially be contradic-
tory, the result of the optimization process will not be a single design, but a Pareto
front, which corresponds to the set of designs that represents a trade-off between the
different objectives.
In real industrial applications, one of the hardest constraints that limits the ex-
ploration is determined by the computing resources available for simulation. In this
study, the complete design space consists of 1,161,216 designs by considering all
combinations of the configuration parameters. This space is clearly too large to
be explored exhaustively, making essential to perform a well defined exploration
strategy.
In this experiment, the time available for simulation limits the exploration to the
execution of at most 8,134 evaluations of the simulator. This information will guide
the selection of the algorithms for the different exploration phases. The rest of this
section illustrates the design space exploration process, starting with the creation
of the optimization workflow, the definition of the experiments, the optimization
process and the results assessment.

8.2.3.1 The Optimization Workflow

The optimization tool used in this experiments is modeFRONTIER, a multidisci-
plinary multi-objective optimization and design tool which is used world-wide in
many application fields like aerospace, appliances, pharmaceutics, civil engineer-
ing, manufacturing, marine multibody design, crash, structural, vibro-acoustics and
turbo-machinery. In the FP7 MULTICUBE project, modeFRONTIER has been en-
hanced to support categorical discrete domain optimization problems, like typical
problems faced in the SoC domain.
The design space exploration process starts with the definition of the optimization
workflow, which is presented in Fig. 8.1 . The workflow can be graphically defined or
can be built automatically using the XML design space definition file (see Chap. 1).