Methods and Tools for Mapping Process Networks onto Multi-Processor Systems-On-Chip - Signal Processing Systems - page 861

Fig. 3 Application,
architecture, and mapping
design space

Multi-Processor System Design Space
Application Architecture

algorithmic transformations configurable hardware

source transformations programmable hardware

Mapping

binding
scheduling

The approach is to model a system as a network of interacting timed automata and
formally verify its behavior by means of reachability analysis using the Uppaal
model checker [ 26 ] .

A comparison of these performance analysis methods is provided in [ 52 ] . Note that
beside being suited for the analysis of real-time systems, analytic models are often
used as the basis for performing system optimization, such as scheduling parameter
optimization [ 69 ] or robustness optimization [ 25 ] .
Finally, one can observe that none of the methods shown in Fig. 2 can fulfill
all the requirements concerning accuracy, scope, and set-up effort. Therefore,
combinations of the different methods have been proposed: Simulation has been
coupled with native execution on the target platform to reduce simulation time
[ 2 , 40 ] . Different analytic methods have been coupled to broaden the analysis scope
[ 42 , 60 ] . Subsystems in simulation have been replaced by analytic models to reduce
simulation time and eliminate the need to generate a detailed simulation model of
a component [ 43 ] . In these efforts, the modularity of KPNs is often leveraged by
using the FIFO channels as the interface between the different performance analysis
methods.

3.4 Design Space Exploration

Designers of MPSoCs face a large design space due to the combinatorial explosion
related to the available degrees of freedom. At several points in the design flow
and at various levels of abstraction, they need to decide between design alternatives.
Specifically, the design space of MPSoCs can be roughly divided into three domains:
the application design space, the architecture design space, and the mapping design
space. These three domains can be further split up, as shown in Fig. 3 .
Exploration of the application design space can be split up into two main kinds of
transformations, namely algorithmic and source transformations. Algorithmic trans-
formations make explicit the coarse-grained parallelism in a sequential application
by transforming it into a KPN. Given a KPN application, source transformations
split and merge processes to trade-off parallelism and communication overhead.