Communications Network Research Institute

An Investigation of the Feasibility of Passive Monitoring for the Support of Routing in Multi-Channel WLAN Mesh Networks

Chenzhe Zhang

Routing protocols for wireless mesh networks need to be aware of the radio environment in order to select suitable transmission paths that have bandwidth sufficient resources to support high throughputs. Thus the accurate representation of the radio link characteristic as a link cost function is a critical design issue where two techniques tend to be used: active probing or passive monitoring. Active probing can provide much useful information, but suffers from an associated overhead penalty. On the other hand passive monitoring does not impose an overhead penalty, but can be difficult to implement in practice.

Another factor that can reduce the capacity of a mesh network is interference arising from having many wireless nodes operating simultaneously in close proximity to one another on the same radio channel. This problem may be alleviated by using multiple channels, however this further complicates the routing mechanism. Passive monitoring can be of considerable benefit in dynamic multi-radio multi-channel systems as it does not require node synchronization in order to operate. Therefore, this project will investigate the potential of passive monitoring for support of routing in multi-channel WLAN mesh networks.

May 2008

During the last few months, a series of experiments have been conducted to investigate the correlation between passive monitoring and active probing techniques used in WLAN mesh networks (WMNs).


In order to optimize the link performance in WMNs, it is necessary to make use of performance metrics that take account of jitter, packet delivery rate, delay, capacity, bandwidth etc. In general there are two techniques that can be employed to obtain these performance metrics: active probing and passive monitoring. Active probing requires accessing the wireless network and broadcasting probe packets. It has the disadvantage that it generates an overhead. On the contrary, passive monitoring is a technique whereby all network transmissions are passively intercepted and subsequent processing produces the performance metrics. It has the advantage that it does not generate an overhead.

Experimental Setup

Experimental Testbed Soekris Board
Figure 1: Experimental test bed and a photo of Soekris net 4521 board used in this experiment.

The test bed consists of 3 Soekris net 4521 boards (single radio) as shown in the diagram above. Two of the net 4521 boards are configured to communicate in the Ad-Hoc mode, and the third node is used to generate bursty traffic on the network. This is intended to introduce variations in the network load. Finally, a control PC is used to control the operations of all 3 nodes using the SSH protocol.

Experimental Approach and Results

The experimental work consists of 3 steps during which different performance metrics have been gathered using the active probing and passive monitoring techniques, and the correlation between these measurements is calculated and plotted. Each test case is carried out with the nodes in different locations, and the typical duration of each test is between 17 to 24 hours, and a minimum of 30 test cases are conducted for each stage.

During the experiment, packet delivery rate is obtained by using active probing. This is done by sending ICMP broadcast packets from Node A to Node B, and the delivery rate is calculated on the receiving node with a program that had been developed based on Pcap library. On the other hand, the passive monitoring uses the CNRI’s Wireless Resource Monitor (WRM) where the idle bandwidth and RSSI values are considered.

Pearson Correlation Coefficient

To calculate the correlation between the two data sets, the Pearson correlation coefficient is used which is a dimensionless index that ranges from -1.0 to 1.0 inclusive and reflects the extent of the degree of correlation between the two data sets. The closer the coefficient is to -1.0 or 1.0, the higher the correlation is between the two data sets. Mathematically, the Pearson correlation coefficient is given by

Pearson Correlation Coefficient

Step 1

To calculate the correlation between the packet delivery rate and the RSSI value of the node.

Correlation coefficient of RSSI and packet delivery rate over all test cases
Figure 2: Correlation coefficient of RSSI and packet delivery rate over all test cases.

From Figure 2 it indicates that out of all 20 test cases during this step, it can be observed that there is no clear correlation between the packet delivery rate and the RSSI value at the receiver nodes.

Step 2

To calculate the correlation between the packet delivery rate and the idle bandwidth (BWidle).

Correlation coefficient of RSSI and packet delivery rate over all test cases
Figure 3: Correlation coefficient of BWidle and packet delivery rate over all test cases.

As shown in Figure 3, there is no discernible correlation between the packet delivery rate and the idle bandwidth for the 20 test cases considered. From these first two steps, it is believed that a single performance metric obtained from passive monitoring is insufficient to establish a correlation with those from active probing.

Step 3

In order to more accurately reflect the link performance observed by passive monitoring technique, the RSSI value and the BWidle value are combined and normalised as follows

The reason the number 30 is used to normalize the function is that in the Madwifi driver used in these experiments, it specifies that the RSSI of a good signal source will have a maximum value of 30.

Correlation coefficient
Figure 4: Correlation coefficient of the combined function and packet delivery rate over all test cases.

As shown in Figure 4, a high degree of correlation can be observed with a correlation coefficient between 0.7 to 0.9. This indicates that passive monitoring can have a good correlation with active probing if the combined function is used for the passive monitoring measurements.

To provide a better graphical insight into this result, see Figure 5 below:

Variation in the combined function
Figure 5: Variation in the combined function value and packet delivery rate over time.

The variation in the network load is produced by the third node is the experiment producing a steam of bursty traffic. Figure 5 above shows how both the packet delivery rate and the combined function react to this variation over time. It can be observed that their variations are quite similar.


  1. We have observed a high degree of correlation between the combined function (i.e. passive monitoring) and packet delivery rate (i.e. active probing).
  2. Active Probing and Passive Monitoring can be used interchangeably to measure link performance. For example, active probing can be used when the network load is low and passive monitoring when the network load is high.

Future Work

October 2008

Continuing on from the previous experiments investigating the use of passive monitoring in mesh networks, a new test bed has been developed and a series of tests have been conducted to study the performance characteristics of multi-radio mesh networks.


In recent years, the shift from “single hop” wireless networks to “multi hop” wireless networks leads to many possible design choices for the architecture of a mesh node. Each mesh node can simply contain a single radio which is used to forward on to a neighbouring mesh node any traffic not intended for it. Alternatively, it can contain multiple radios in order to support a number of concurrent wireless point to point links with neighbouring nodes. When these radios are operated on non-overlapping frequencies, theoretically the capacity of the network increases as a function of the number of radios. However, there are many constraints and obstacles associated with these multi-radio mesh networks such as radio leakage, cross talk which tend to limit the maximum throughput that can be achieved. The evaluation of the potential of these multi-radio mesh networks have typically been addressed through simulations. Instead, this project adopts an experimental approach to investigate the factors that limit throughput performance.

Experimental Setup

Experimental Setup
Figure1: Experimental test bed used in this experiment.

The experimental test bed comprises 2 PCs that have up to 4 WLAN NICs in each. These 8 NICs have been configured to operate in the Adhoc (i.e. infrastructureless) mode to form 4 pairs operating on different channels. The NICs operate in the 802.11a mode with the antenna diversity feature disabled and a fixed transmission rate of 54Mbps on each interface.

Experimental Approach

  1. Initially, a baseline performance is established by measuring the maximum achievable throughput between a single Tx/Rx pair (i.e. involving just 2 NICs). The other 6 NICs are disabled.
  2. Having established a stable and reproducible baseline, the impact of multiple wireless interfaces within a node is studied. The intention is to determine if there is any negative impact on throughput performance as a result of cross-talk between the wireless interfaces. During this study, a single receiving node with a single interface is used and the number of transmitting interfaces in the transmitting node is increased until all four PCI slots are occupied. Each time an additional wireless interface card is inserted, the card is set to the monitor (passive) mode and tuned to an orthogonal channel. The maximum achievable throughput is measured as described in step 1 above. The CPU usage is also noted.
  3. Steps 1 and 2 are then repeated with a different tool (i.e. D-ITG) to generate a saturated network condition, where the additional cards are set to the AP mode. The throughput results for the saturated network are recorded.

Tools and Commands

  1. Throughput statistics under non-saturation are obtained by using Iperf: Iperf client is setup on the TX NICs and Iperf server is configured on the RX NICs. UDP traffic is used throughout all the experiments. The duration of the test measurement is 30 seconds and each measurement is repeated 3 times and the average is taken.
  2. Throughput statistics under saturation are obtained by using D-ITG tools: DITG send is configured on TX NICs and DITG recv is setup on RX NICs. Again UDP traffic is generated in order to saturate the network. The test duration is 30 seconds and each measurement is repeated 3 times and the average is taken.

Results and Explanation

A: Throughput statistics under non-saturation condition at the TX node.
Table 1
Table 1: Throughput under non-saturation at the TX node with different packet sizes and number of active interfaces.

Figure 2
Figure 2: The variation in throughput under non-saturation conditions with multiple wireless cards.

Figure 3
Figure 3: Effect of additional NICs on throughput under non-saturation with a packet size of 1470 bytes.

Table 1 and figures 2 and 3 indicate that when the network is not saturated and provided that there is sufficient CPU processing power, the capacity of a network decreases as a function of the number of radios. As expected the throughput increases when using larger packet sizes in traffic owing to more efficient use of the transmission medium. It should be emphasized that each additional NIC card has been placed in a monitor (passive) state, and therefore each additional NIC imposes an additional load on the CPU as every frame on the medium is monitored by these additional NICs and hence requires servicing by the CPU. This ultimately leads to a decrease in the throughput performance. This is confirmed by the results in section B below.

B: Throughput statistics under saturation condition at the TX node.
Figure 4
Figure 4: Throughput statistics under saturation condition with a reference to CPU idle% @ 1470 bytes.

Figure 5
Figure 5: Throughput comparison in 2 different modes under saturation @ 1470 bytes.

During this phase of the experiment, the D-ITG tool sets are used to generate large amounts of UDP traffic in order to fully saturate the network and the CPU. The packets are 1400 bytes in size and are created at the rate of 5000 every single second. From the data analysis carried out after the experiments, it is observed that CPU idle% < 1% throughout the tests. Furthermore, Figure 4 shows that the throughput does not vary even when additional NICs are inserted. This result fully confirms the observation made in Section A that the decrease in throughput is primarily due to the finite processing power, i.e. crosstalk does not appear to be a limiting in determining the maximum achievable throughput. Figure 5 shows the throughput comparison when the additional NICs are set to the Monitor mode and the AP mode under saturation conditions. There is a slight increase in throughput when using the AP mode which is due to the fact that AP mode does not require that much CPU resources when compared to the Monitor mode.

December 2008

Having studied the impact of multiple NICs to the throughput performance at the TX side, a new series of tests have been carried out to investigate the impact at the RX side. The same testbed is used as described in the previous section, see Figure 1 above.

Experimental Approach

Similar procedures are applied at the RX node, but here different operation modes of the NICs are used while the saturated throughput is recorded. This is to examine the impact of the various operation modes on the network performance. Also the TX transmitting power is varied to study its impact on throughput. During each test, link quality is measured and recorded using the Madwifi driver. The D-ITG tool sets are used to generate a saturated network condition and to record throughput for each scenario.

Results and Explanation

Table 1
Table 2: Results recorded at the RX node under different test scenarios.

From Table 2 above, several observations can be made as follows:

  1. Both the signal level and noise level decreases when the number of active NICs at the RX node increases.
  2. When the additional NICs are placed in inactive mode, it can be observed that the signal level increases as well as the link quality plus a slight increase in the throughput.
  3. There is no significant change in the throughput when additional NICs are inserted into the RX node.
  4. There is no major difference in throughput between various operation modes.
  5. A significant decrease in throughput can be observed when the TX transmitting power is set to minimal.
  6. Throughput changes in relation to link quality.

Future Areas of Study