Information Technology Reference
In-Depth Information
result in optimal speedups. Multi-core processors and distributed
systems are suciently different (e. g., regarding bandwidth and com-
munication reliability between the cores respectively processors) to
not just carry over well-known strategies for parallelization. Instead,
to exploit the full potential of multi-core processors their specific
characteristicshavetobetakenintoaccount.TheGRAMSreference
model is independent of the implementation of different simulation
engines which may exploit the underlying hardware in an optimal
fashion. Therefore, the GRAMS reference model provides also the
necessary foundation for fulfilling the second requirement.
3.3 Existing approaches and their limitations
Parallel simulation may either decompose the computation on a spatial
basis or on a temporal basis [36, p. 177ff.]. Time-parallel simulation
(i. e., a simulation using temporal decomposition) is highly sensitive
to the state-matching problem [36, p. 178f.] and is considered hard
to employ in case of complex state transitions [67]. As agent-based
models are most often used to represent complex systems with vari-
ous interactions between the agents, time-parallel simulation is not
eciently applicable to this kind of models.
This assumption is reinforced by the fact that agent-based models
usually have a strong emphasis on the simulated environment (cp.
Chapter 2.4). Lees et al. discuss the importance of the environment
primarily with respect to distributed simulation and point out that the
agents' environment resembles a large shared state of a multi-agent
simulation. Therefore, 'the ecient simulation of the environment of
a multi-agent simulation is perhaps the key problem in the distributed
simulation' [76]. This statement is in accordance with findings of
Logan and Theodoropoulos [127, 78, 128].
Various approaches for parallel multi-agent simulation have been
proposed. Notably, all approaches known so far consider only spatial
decomposition. Also, various simulation toolkits support parallel
execution. A good overview of existing tools for parallel agent-based
