Information Technology Reference
In-Depth Information
A highly scalable agent-based model for simulating the spreading
of diseases is presented in [97]. As the author points out, the most
immediate problem is that the size of the population (in this case
approximately 300 million agents) exceeds the memory of a single com-
puter. Distribution of the agents is therefore driven by the constrained
resources of a single computer (especially memory limitations). The
devised load-balancing approach tries to weigh the influence of smaller
agent subsets using a larger number of computers versus the increased
communication overhead introduced with each new agent subset. In
this special case, minimizing inter-node communication is the way to
achieving highest speedups. However, the results presented make use
of only two nodes, each of them executing eight threads.
Regarding existing multi-agent toolkits, various attempts have been
made regarding distributed multi-agent simulation. Distributed exe-
cution based on the Java Agent Development Framework (JADE) is
discussed in [86]. As JADE is primarily designed to be a framework
for multi-agent systems, it lacks several features usually regarded
essential for multi-agent simulation (e. g., proper time management).
Furthermore, frameworks originating from the domain of multi-agent
systems are usually designed to execute a small number of agents. The
major findings presented in [86] are that inter-agent communication
may cause substantial delays and that (depending on the number of
agents) local or global synchronization should be preferred. Redu-
cing communication as well as redistributing agents is considered for
improving the performance.
Similarly, Gianni et al. propose DisSimJADE as a distributed
simulation framework on top of JADE [44]. Following the layered
SimArch architecture [43], DisSimJADE allows transparent execution
either in a local or distributed environment. A brief example scenario
is provided, using a graph-based environment to represent a manu-
facturing system. Actual results and experiences are not reported
in [44].
A distributed discrete-event simulation engine that explicitly high-
lights the isolation of the modeling domain from the simulation engine
is described in [104]. Separating these two aspects is stressed as an
