Performance Improvements for Large-Scale Traffic Simulation in MATSim - Computational Approaches for Urban Environments - page 215

9.3.3.1 Implementation

The parallelized version of event handling is called parallel event handling and
allows the user to specify how many threads should be dedicated to this module
during the simulation. The current implementation applies a round-robin approach
(Hahne 1991 ) to assign event handlers to threads. This means it tries to assign the
same number of event handlers to each thread. It is obvious from Fig. 9.3 that this
approach is suboptimal because it would be best to put the event handler writing out
events to a separate thread and the other handlers on a second one.
Fortunately, writing out events to a file is not part of the communication interface
between JDEQSim and MATSim, which is different in DEQSim. This means the
event file is needed only in the last iteration as a backup for further analysis.
Additionally, events can be written out at predefined intervals, e.g., each 10th
iteration, in order to analyze intermediate results of a bigger simulation while it is
still running. This means that during most iterations, a quite balanced parallelization
of event handling is possible using up to four cores if considering the default
simulation without any other scenario-specific event handlers.
While the implementation of parallel event handling could be improved in several
ways, as discussed in the future work section, the reason to start with a simple
implementation is that the existing interfaces did not need to be changed.

9.4 Experiments and Results

9.4.1 Single Thread JDEQSim with Parallel Event Handling

While the parallel version of JDEQSim is still experimental work, the nonparallel
version of JDEQSim is planned to be run with parallel event handling in the near fu-
ture for large-scale scenarios. 1 Therefore, experiments presented here are conducted
in the latter configuration to demonstrate the performance gains when comparing it
to the current state-of-the-art micro-simulation in Java (JQueueSim). Furthermore,
the performance gains due to parallel event handling are also demonstrated. While
the focus is on the comparison of JDEQSim with JQueueSim, runs with DEQSim
are also conducted for reference.

9.4.1.1 Scenario Setup and Hardware

The simulations are conducted on a NAVTEQ road network (NAVTEQ 2009 )
for Switzerland, with around 882 K links. A population sample of the people
surrounding the city of Zurich who drive cars is used, containing around 614 K

1 Have been conducted in 2010 (see, Meister et al. 2010 )