Simulation

There are two main issues when doing simulations except for modeling of the physical environment. The first one is the modeling of how the nodes move in the mobile ad hoc network. This is very tricky if you don't know how the nodes will move in the real world. There are a number of suggestions how to do this and some are found in the next section.

The other issue is that in order to compare two different networks and describe their behaviour you need some kind of metrics. If you are trying to evaluate two different routing protocols you may want to know the mobility, i.e how hard it is to handle routing in the network. It is a popular term and in this case it merely describes the activity of link-changes caused by external physical interference.

Mobility models used in simulations

Random waypoint model (RWM)

Johnson and Maltz describe in [110] the Random waypoint model. It works as follows. All nodes are uniformly distributed around the simulation area at starting time. Each node then choose a random destination and moves there with a speed uniformly distributed over . Then there is a pause time which could be selected be 0 to give continuous motion.

Random direction model (RDM)

In [95] Royer et al describe another random based model. This a more ``stable'' model than a random waypoint model. At start the nodes selects a random direction and starts to move along it. Since the area of simulation is confined the node may end up reaching one of the boundaries during the simulation. When a boundary is reached the node pause for a given time and then chooses a new direction to travel. Since the node is on a boundary the selectable angle is 180 degrees. The result of this model is a more stable distribution of the nodes than the RWM (). The behaviour can be thought as a micro-cell of a larger area which is a useful property.

Modified Random direction model (MRDM)

A second more advanced version described by Royer et al in  [95]. To give a even more realistic simulation the Random Direction Model () was extended with a extra choice for the nodes when their pause time is over. The nodes don't have to travel all the way to the boundary but could stop anywhere along the path.

Brownian model (BM)

Hu and Johnson describe in [100] another way of modelling the speed of the nodes. Changes speed and direction at discrete time intervals and at the beginning of each interval each node chooses and moves with velocity vector . This model is very similar to the random direction model except for the speed which is smooth in this model.

Column model (CM)

A mobility model suited for experiments is described by Sanchez in [112]. Nodes are only moving along the x-axis. The initial position of node is and the node changes the speed at the discrete intervals. This will produce a mobility pattern that is one dimension simpler than the random mobility model since the nodes only move along the x-axis.

Random Gauss-Markov model (RGM)

Uses discrete time intervals to divide up the motion. The nodes update their velocity vectors at the beginning of each interval according to:

is a random variable with mean 0 and variance . This model is describe by Sanchez [112] and further developed by Liang and Haas [109].

Pursue model (PM)

Another model done by Sanchez [111] in order to try to create group movement. One node in each group is moving according to the random waypoint model (). The rest of the group is moving towards the target that the ``leading'' is aiming for. The speed of the pursuing nodes is chosen uniform random in the range .

Exponential Correlated Random model (ECR)

The ECR is able to model all possible movements of individuals and groups. This is done by changing the parameters of a motion function. A new position is a function of the previous position b to which a random deviation is added. The function can be defined either for a single node or a group at time t. r is a random Gaussian variable with variance . The parameters are then changed to give different mobility patterns. Very hard to create a predefined motion pattern by selecting the parameters. This model is described by BBN in [115].

Reference Point Group mobility model (RPGM)

Ho et al describes another way to simulate group behaviour in  [96] where each node belong to a group where every node follow a logical centre reference point. The nodes in a group are usually randomly distributed around the reference point. The different nodes use their own mobility model and is then added to the reference point which drives them in the direction of the group.

This general description of group mobility can be used to create a variety of models for different kinds of mobility applications.

Individual Simulated Behavioral model (ISB)

This is another new and different idea how to do more accurate and better simulations. They use a theory about an individually simulated behavioral model where all objects has their own properties. They verified their idea with DSR and proved that it generates reproducible and ``realistic'' mobility patterns. [121]

Mobility metrics

Geometric-based mobility metric

Johansson et al [108] described a geometric mobility metric as

where is the physical position of node k at time t, T is the length of the test and N is the number of nodes participating in the test. The sum is calculated over all node pairs over the scenario duration. Hu and Johnson [93] used exact this model in their simulations, while Johansson et al [108] used an approximation of this geometric mobility metric in their simulations.

Minimal route-change metric

This metric described by Hu & Johnson [100] basically calculates the link-breaks in a route counted by the metric. It is assumed that all link are bidirectional. The routes counted by the metric can either be the ones between the nodes that communicate over all pair of nodes regardless of the traffic during the measured time.

Discussion of simulation models and metrics

Mobility models not accurate enough

Simulations have shown that the existing models are not accurate enough for real world simulations. In [96] results show that real ad hoc environment cannot be simulated with a ``random walk'' type of mobility model. The more advanced models that exists provide a much better simulation than the random walk types but have the disadvantage that they demand much more computation time. Models like the ECR have great possibilities but have shown very hard to control to get the desired motion patterns.

Mobility models not used combined

A common view of an ad hoc network is that nodes join/part/move in the network relatively often. The result of this is that the traffic patterns get more important since an mobile/unstable network is less suitable for some types of traffic, e.g. high-speed backbone traffic with demands on low latency. The most effective traffic pattern for this kind of network seem to be bursty low-volume traffic e.g. www/mail. The reason for this is the low demand of long-lived routes and latency.

As mentioned, the movements of the nodes in an ad hoc network is very essential for the routing protocol and therefore the mobility model is important for simulation results. However, all the simulations studied have only used one of the mobility models and that fact illuminate a problem that should be investigated. To conclude, there's a need to look carefully at the mobility models used in simulations as these are crucial for the real-world functioning of an hoc network.

Metrics not advanced enough to capture needed properties

Johansson et al [108] used the protocols DSDV [19], AODV [4] and DSR [41] to simulate a conference, a event coverage and a disaster area. The conclusion was that the metric worked fine when doing simulations.

Whereas the geometric mobility metric give us a view of how mobile the nodes are physically the minimal route-change metric give us a view how stable the links in the network are. Thus should these two combined give a quite good view of how the network behave. High mobility but at the same time low route-change could describe a network where the almost all nodes move, but not enough to potentially affect the routes. If it was possible to do accurate positioning with our testbed this would be a hypothesis to verify since the propagation in the real world is very unstable.

But since we have no possibilities to do accurate positioning in large scale experiments due to the fact that it would too expensive there is a need for new metrics. These metrics can not as stated rely on expensive positioning devices and should use the data available about the ongoing traffic for best economics.

The propagation models used are too simple

The simulations that have been done have been using a very simplified model of the real world environment. Most of the simulations have been using NS-2 [76] which has an extension [77] to simulate a wireless environment. The extension uses the Free Space Path Loss (FSPL) [97] model or Two Ray Ground Reflection (TRGR)[97] depending on the distance:

Where is the distance, the wavelength and the the heights of the transmitter and receiver antennas.

It can not simulate objects as walls etc that have influence on the propagation. In [108] they extended NS to be able to simulate solid objects that block the signal if LOS is lost. This is still a very simple approximation of the complex real world. A more accurate model to do more accurate simulations is the Simulation of Indoor Radio Channel Impulse-Response Model (SIRCIM) [78] which simulates fading, barriers, foliages, multi-path interference etc. It has been implemented by  [105] in Global Mobile Simulation library GloMoSim  [79] which uses PARSEC [99]. The computation time when using this model is about three magnitudes longer and it has to be tuned to each specific scenario. The model have not been used in any simulation results but when doing simulation of complex indoor environments models like the SIRCIM are required.

Frequency related modelling is required

Even more advanced model have to be constructed in order to be able to simulate ``soft'' objects like people since they have great influence on the propagation, especially in the 2.4 GHz band. When a person is standing in the LOS between two nodes equipped with WaveLAN 2.4GHz cards the signal is dampened with about 15 dBm which is equal about 20 meters in distance according to [120]. This fact is critical if you try to do positioning systems using only the signal level.

This also give a hint that propagation models for objects that absorb radiation will be required to do simulation of scenarios like a conference where a lot of people will be moving around. To be able to create these models there is a need for studies of the absorbation of water-rich substance around the free 2.4GHz band. For example, if there are a lot of humans in between it might help to change the channel to a higher/lower frequency.