Information Technology Reference
In-Depth Information
Additionally to the set of agents, the environment is also split up
and distributed onto available nodes.
Obviously, decomposing and distributing the simulated environment
onto multiple computing nodes allows simulation of very large models.
This is especially important if a model is too large to be executed on
a single computing node.
The decomposition approaches described for node-level paralleliza-
tion can also be applied to cluster-level parallelization. Four different
combinations are distinguished:
Static/ Agents only
The environment is replicated onto all nodes. The static partitioning
of the set of agents is uninformed as no specific information is taken
into account (cp. static partitioning on node-level as described in
Section 8.1.3).
Static/ Agents and environment
Depending on the type of environment, a decomposition into
n
partitions is computed. The set of agents is distributed onto the
nodes according to the environmental decomposition. Both decom-
positions of environment and the set of agents are computed only
once (static).
Dynamic/ Agents only
The environment is replicated onto all nodes. A dynamic parti-
tioning of the agents may take additional information gathered
during runtime into account, e.g., interaction pattern (cp. dynamic
partitioning on node-level as described in Section 8.1.4).
Dynamic/ Agents and environment
Besides only taking into account additional information for com-
puting the partitioning of the agents, the additional information
gathered during a simulation is also taken into account to compute
a decomposition of the environment. As described in Section 8.1.4
the benefits of a better decomposition have to be balanced against
the resources required for computing and updating the partitioning.
