This paper explains a parallelization method for the wide area micro traffic simulations. These kind of simulations should be parallelized in order to achieve an efficient use in terms of computing resources. Our approach here is to run such a simulation on a cluster of PCs which is much more affordable than to buy a parallel computer.
Parallel computing comes with some important issues such as domain decomposition, data sharing/exchanging and communication between processors. We represent our approaches on these issues which will affect the performance of a parallel system.
A well-behaved parallel system is load balanced. In order to achieve load balancing, one should be careful with the domain decomposition. If the parallel application does not have constant loads on the processors, it is better to use a dynamic/adaptive method to disaggregate the domain onto processors.
Data sharing among processors can be employed by using either shared-address space method or message passing approach. Message passing is more efficient in terms of bandwidth and memory usage. Each processor is independent but at the same time in a cooperation with the other processors when necessary. As the name implies, the communication is done through the messages exchanged among processors.
We finally demonstrate how computing time for a parallel traffic micro-simulation can be systematically predicted. An important result is that a typical city with 20024 links network runs about 40 times faster than real time on 16 500 MHz Pentium computers connected via switched 100 Mbit Ethernet. These are regular desktop/LAN technologies. When using the next generation of communications technology, i.e. Gbit Ethernet, we predict the same computing speed for a much larger network of 200000 links with 64 CPUs.