...P

nagel@inf.ethz.ch. Postal: ETH Zentrum IFW B27.1, 8092 Zürich, Switzerland

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...

marcus.rickert@topmail.de

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... congestion.

It is sometimes argued that TRANSIMS is unnecessarily realistic for the questions it is supposed to answer. Although we tend to share the same intuition (see, for example, our work on the so-called queue model [39]), we think that this needs to be evaluated systematically. We also expect that the answer will depend on the precise question: It will be possible to answer certain questions with very simple models, while other questions may need much more realistic models.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...DYNEMO,DYNEMO:parallel,

DYNEMO is not strictly a micro-simulation - it has individual travelers but uses a macroscopic approach for the speed calculation. It is mentioned here because of the parallelization effort.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... TRANSIMS-1999

There are two versions of TRANSIMS with the number ``1.0'': One from 1997, ``TRANSIMS Release 1.0'' [5], which we will refer to as ``TRANSIMS-1997'', and one from 1999, ``TRANSIMS-LANL-1.0'' [41], which we will refer to as ``TRANSIMS-1999''. From 1997 to 1999, many features were added, such as public transit with a different driving logic, or the option of using continuous corrections to the cellular structure. For the purposes of this paper, the differences are not too important, except that computational performance was also considerably improved.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...fig:parallel):

Instead of ``split links'', the terms ``boundary links'', ``shared links'', or ``distributed links'' are sometimes used. As is well known, some people use ``edge'' instead of ``link''.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... CPUs.

For simplicity, we do not differentiate between CPUs and computational nodes. Computational nodes can have more than one CPU -- an example is a network of coupled PCs where each PC has Dual CPUs.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Mbit/s.

Our measurements have consistently shown that node bandwidths are lower than network bandwidths. Even CISCO itself specifies 148000 packets/sec, which translates to about 75 Mbit/sec, for the 100 Mbit switch that we use.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... small.

An event-driven simulation could be a counter-example: Depending on the implementation, it could be extremely fast on a single CPU up to medium problem sizes, but slow on a parallel machine.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... CPUs.

This is possible because of the specific purpose TRANSIMS is designed for. In real time applications, where absolute speed between request and response matters, the situation is different [8].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.