Performance Improvements for Large-Scale Traffic Simulation in MATSim - Computational Approaches for Urban Environments - page 205

9.1 Introduction

Traffic simulations can be performed at different levels of detail. One common
technique is to model traffic as flows consisting of aggregated number of cars
between different areas (de Dios Ortúzar and Willumsen 2011 ). While this technique
is quite fast, it does not allow modeling of individual preferences or temporally and
spatially detailed analysis.
In contrast, agent-based travel demand models like MATSim represent each
person in the simulation as an individual agent (MATSim 2009 ). The travel demand
by each agent in this case is based on an activity-based model, where activity
times and durations are time dependent (Axhausen and Gärling 1992 ). Furthermore,
a dynamic micro-simulation model is used to model detailed traffic interactions,
which are again time dependent.
This enables various kinds of new applications, which are not possible with the
first approach - e.g., detailed modeling of car sharing (Ciari et al. 2008 ) or modeling
the charging behavior of electric vehicles (Waraich et al. 2009 ). Also, commercial
applications of such detailed micro-simulations can be envisioned. For example,
companies owning advertising space could offer a more sophisticated service to
customers, where not only the traffic volume along a road determines the price but
also the target audience of an advertisement is considered.
While more powerful, such detailed micro-simulation models are more expen-
sive, in terms of computation time, than aggregated models. This chapter describes
efforts to improve the performance of an agent-based micro-simulation model called
Multi-Agent Transport Simulation Toolkit (MATSim 2009 ). This model is aimed at
the simulation of large travel demand scenarios. But in order to perform a simulation
of a full population run of Switzerland, with 7.3 million agents on a high-resolution
navigation network, it is estimated that the existing Java-based micro-simulation
in MATSim would require around 3-4 weeks. In the direction of reducing the
computational time of MATSim, two of its central components are redesigned.
The first component is the mobility simulation where the traffic dynamics are
modeled. In order to make the mobility simulation faster, a new micro-simulation
is implemented based on the ideas of an existing event-based micro-simulation
(Charypar et al. 2007a ). While multiple distributed and parallel traffic simulations in
the C CC programming language have been implemented in the past (Barceló et al.
1998 ; Nökel and Schmidt 2002 ; Nagel and Rickert 2001 ), to the best of the authors'
knowledge, this chapter presents the first large-scale implementation of such a
simulation in the Java programming language (implemented mid-2009). Therefore,
this chapter also discusses specific challenges for large-scale traffic simulation in
Java, which has not been discussed in the related C CC literature.
A second major performance improvement achieved in this work is related to an-
other core component of the MATSim framework called event handling. This com-
ponent is needed to process the results of the mobility system and is therefore essen-
tial for integration with other MATSim internal components and also for extension
of the MATSim model. Parallel computing is used to make event handling faster.