Information Technology Reference
In-Depth Information
- The obvious choice is to partition on a per-agent -basis. This way,
all issues related to the simulation of one agent are concentrated
within one partition and subsequently within one thread. This
thread in turn may be responsible for multiple agents, of course.
- A second possibility is to partition on a per-action -basis. This
approach does not consider agents as a whole, but treats sensor
and effector actions of all agents individually. Actions which do
not interfere with each other may be executed concurrently. The
main problem of this approach is that access to shared memory
of a single agent may be distributed across multiple partitions
(and subsequently across multiple threads).
Many more partitioning strategies may be thought of, yet these are
the basic strategies. The crucial aspect of all partitioning strategies is
the identification of actions or agents which are independent of each
other and may therefore be executed concurrently.
Despite the partitioning strategy ( How? ) the partition computa-
tion strategy ( When? ) or update rate of a partitioning has to be
considered:
A static partitioning is computed once at the beginning of a simu-
lation and only updated when necessary (e.g., new agents joining a
simulation). The main benefit of this computation strategy is its
simplicity, which on the other side may lead to suboptimal parallel-
ization. This is due to at least two reasons: An ideal partitioning
may not be known in advance (i. e., without running a simulation),
and an ideal partitioning may change over time.
A dynamic partitioning is updated regularly (either in fixed time
intervals or due a specific condition). This computation strategy
may take into account much more information than a static compu-
tation strategy. Based on statistic information gathered during a
simulation, an ideal partitioning may be determined. Also, changes
in a model during simulation (like agents joining and leaving a
model) may be taken into account. A drawback of this strategy
is the considerably more complex implementation. Furthermore,
Search WWH ::




Custom Search