Information Technology Reference
In-Depth Information
be computed for a partitioning do not predict the actual reduction of
runtime.
Choosing an appropriate indicator to deduce runtime improvements
is usually only possible if the currently executed model and its behavior
are well-known. Based on in-depth knowledge about a model and its
peculiarities, a well-suited indicator may be chosen. This proceeding
requires lots of knowledge about a model itself as well as detailed
knowledge about possible indicators. Bearing in mind that a model
developer should not need in-depth knowledge about the simulation
engine actually executing a model, this approach seems to be not
feasible. Ideally, a simulation engine takes care of all aspects related
with model execution, especially of executing a model in a most ecent
way.
One possible approach is to apply methods and techniques of self-
optimization. A simulation engine could use an uninformed partition-
ing strategy at the beginning and regularly update the partitioning.
Of course, the easiest realization of self-optimization would use the
indicators mentioned before (e.g., minimal network tra
c). A more
sophisticated approach could employ methods of genetic algorithms
to create and update a partitioning, combined with an evaluation of
the current partition based on the actual runtime. For this purpose,
a simulation engine could measure the runtime for a while (e.g., 100
time steps), update the partitioning and continue with the simulation.
Provided that a simulation is long enough, a good partitioning may
be found after a while.
8.1.5 Partitioning strategies on cluster-level
Node-level parallelization assumes that it is possible to execute a model
on a single computing node. In contrast, cluster-level parallelization
allows to split and distribute a model across multiple computing nodes.
Similar to node-level parallelization two approaches are possible:
Only the set of agents is distributed onto available nodes. The
environment is replicated on each node, i. e., each node contains
information about the whole environment.
