Model partitioning and multi-level parallelization - Modeling and Simulation of Complex Systems - page 124

Regarding multi-level parallelization each of the approaches applied
on cluster-level may be combined with the approaches for node-level
parallelization.
8.2 Parallel execution
As illustrated in Figure 8.1 parallel execution of an agent-based simu-
lation is possible on different levels. Within this thesis three levels of
parallelization are distinguished:
Parallelization on cluster-level
Parallelization on node-level
Parallelization on processor-level
Although parallelization can be employed on these three levels, it is
important to keep in mind that there is exactly one agent-based model
which is executed. The execution is performed by a simulation engine
which may be implemented in any fashion (e.g., as a distributed
system).
Parallelization on cluster-level refers to executing a simulation en-
gine on multiple interconnected computing nodes. Each component of
a simulation engine running on a dedicated computing node may be
parallelized further. Node-level parallelization refers to parallelizing
the execution of a simulation engine on a single node, i. e., to use mul-
tiple processors of a computing node. Processor-level parallelization
addresses the issue that each processor may contain multiple cores
and therefore may execute multiple threads in parallel.
The combined application of parallelization strategies on different
levels is called Multi-Level Parallelization . Similar terms which may be
found in literature are nested parallelism [117]aswellashierarchical
parallelism and multi-stage parallelism [82]. In the following, only the
term parallelization is used.
As parallelization issues on processor- and node-level are very sim-
ilar, these two levels are not treated independently. The two resulting