Information Technology Reference
In-Depth Information
oration between model developers and developers involved in providing
the necessary development and runtime infrastructure has to be con-
sidered. In the simplest case, the infrastructure consists only of
the software required for executing a model. Issues like providing
high-performance computing capabilities and distributed (grid-like)
execution have to be covered for effectively simulating large-scale
models.
Effective model execution
The shift towards microscopic modeling is closely connected with the
availability of the required computing power. While the increase in
computing power follows Moore's Law for more than fourty years,
software development does not nearly keep pace. The computing power
provided by hardware and the portion effectively used by software
are diverging. This is known as
eciency and programmability gap
,
and exactly this gap is becoming one of the major issues in software
development nowadays [31,p.7],[87], [3, p. 7f.]. This trend is even
accelerated and amplified by the development of multi-core processors
which will by far dominate future processor designs [37, 126, 64, 2].
The challenge of bridging this gap obviously applies similarly to
simulation. Given the demand of more and more complex models, it
is necessary to exploit the computing power of multi-core processors
in an optimal fashion [22, 26]. On a conceptual level, agent-based
modeling seems to be a very promising approach to do so. As all
agents act in parallel, this should allow for a high degree of paral-
lelization. Furthermore, agent-based models offer a high scalability
both in number of agents and level of detail within single agents. Yet
besides missing foundations in the area of agent-based modeling in
general, concepts and procedures as well as actual technical solutions
for exploiting model-inherent parallelism are also missing. Last but
not least, considering large-scale models and their collaborative devel-
opment, it should be possible to develop and test models locally while
executing them on high-performance computers afterwards [125].
