Performance Improvements for Large-Scale Traffic Simulation in MATSim - Computational Approaches for Urban Environments - page 218

1
b C

S.n/ D (9.1)

1b
n

To give an example of Amdahl's law, if 5 % of a program cannot be parallelized,
then it is not possible to achieve a speedup of more than 20. The implication of
Amdahl's law is present both for DEQSim and JDEQSim runs but at different
points. For DEQSim, the interface between the micro-simulation and MATSim is the
bottleneck. Because of the I/O overhead of the communication between the micro-
simulation and MATSim, a speedup of even two seems impossible. This means that
more than 50 % of the micro-simulation consists of parts which have to be executed
sequentially. A second and smaller part of nonparallelizable code present in the
DEQSim runs is the event handling, which cannot run in parallel mode for DEQSim
at the moment. In case of parallel event handling, the maximum achievable speedup
is limited by the slowest event handler.
This first experiment suggests that to make most efficient use of CPUs, JDEQSim
should be run with one parallel event-handling thread. As the machine used in
this experiment has around 128 GB RAM and 16 cores and the scenario uses less
than 15 GB of RAM, several JDEQSim runs could run in parallel, which is useful
especially during the calibration phase.

9.4.2 Influence of Network Size

In the previous experiment, JDEQSim performed around four times faster than
JQueueSim. But this cannot be generalized, because if the network is congested,
then JDEQSim can be much faster than JQueueSim. Such congestion can happen
especially during the initial iterations, in which the routes are far from optimal and
can lead to a simulation period stretching far beyond 24 h. This can lead to long run
times for JQueueSim as its runtime is directly correlated to the simulation period.
Furthermore, different ratios of network size and population can widen the gap
between the speedup of JQueueSim and JDEQSim, which is highlighted here. In this
experiment, all micro-simulations are run using two threads. Both JDEQSim and
JQueueSim runs are performed with parallel event handling, using a single thread
and no events are written to the hard drive. The network capacity is chosen in such a
way that no congestion should happen in order to remove possible adverse influence
of this on JQueueSim. The three scenarios which are considered are:

￿ Scenario A: Network with 882 K links and 61 K agents (36 M events)

￿ Scenario B: Network with 61 K links and 616 K agents (58 M events)

￿ Scenario C: Network with 882 K links and 614 K agents (363 M events)

The results of the experiments in Fig. 9.5 show that DEQSim and JDEQSim scale
linearly with the number of events. Only in Scenario A, in the case of DEQSim, the
I/O overhead of loading the network is immense compared to the actual simulation
time.