Methods and Tools for Mapping Process Networks onto Multi-Processor Systems-On-Chip - Signal Processing Systems - page 871

The approach taken in DOL relies on using compositional performance analysis
where a system is modeled as a set of processing components that interact via
event streams. This is a good match for MPSoCs as well as for KPNs. Contrary
to other approaches, the approach is rather flexible in that it is not limited to a
certain system architecture, scheduling policy, or execution semantics. Specifically,
Modular Performance Analysis (MPA) is integrated into the DOL design flow. In the
following, the basic concepts of MPA are reviewed. Afterwards, it is summarized
how MPA is integrated into the DOL design flow.

4.4.1 Modular Performance Analysis

Modular Performance Analysis (MPA) [ 8 , 71 ] is an analytical approach for the
analysis of real-time systems. It is based on real-time calculus [ 64 ] which has its
foundations in network calculus, a method for worst-case analysis of communi-
cation networks [ 9 , 44 ] . With MPA, hard upper and lower bound for performance
metrics of a distributed real-time system can be computed. As shown in Fig. 9 , the
performance model of a system is decomposed into a network of abstract processing
components that model the computation and communication in a system. These
processing components are connected by abstract event streams that model the
timing properties of the data streams flowing through the system. Finally, resources
are modeled by resource streams that model the availability of processing resources
to computation and communication tasks. Different scheduling policies can be
modeled by differently connecting processing components and resource streams.
As an example, Fig. 9 illustrates fixed priority (FP) scheduling on processors and
time-division multiple-access (TDMA) scheduling on the bus.
The processing components modeling computation and communication are
characterized by the worst-case/best-case execution demand and the minimal and
maximal token size, respectively. Event streams are characterized by so-called
arrival curves and resource streams by service curves. Summarizing, these abstrac-
tions allow the modeling of computation and communication on heterogeneous
resources in a unified manner.
Based on these abstractions, the system is analyzed by consecutive propagation
of event streams between components. Depending on the mapping of processing
components to a resource and its scheduling/arbitration, the timing properties of the
streams change. System properties such as resource utilization, system throughput,
end-to-end delays, or buffer sizes can be derived this way. Tool support for actually
performing the analysis is provided by a freely available Matlab toolbox [ 70 ] that
implements the underlying algebraic operations.

4.4.2 Integration of MPA into the DOL Design Flow

It has been mentioned that the goal of system synthesis is to bridge the imple-
mentation gap, that is, to refine a high-level system specification into an actual