TMA portal

We have developed a passive bandwidth estimation scheme for 802.11-based wireless networks. Our scheme involves passively monitoring the wireless channel in order to identify the “idle” and “busy” periods. Using this information, we can calculate the channel utilization and hence the available bandwidth.

We use the time-stamp information available from the open-source wireless driver (Madwifi) and use it to infer the times at which a packet was sent out and the time at which the acknowledgment for that packet was received.  Combining this information with the time spent by the node in sensing the channel and other activities (DIFS and SIFS) gives is the net channel utilization. We currently incorporate node back-off by assuming an average value for the back-off. It is not feasible to obtain the accurate value of the time for which a node backs off in present day wireless systems. Eliminating the back-off value from the calculations will provide us with an upper bound on the available bandwidth.

A slight drawback of such an approach is that each node only calculates the channel utilization in its own neighborhood. In order for two nodes to communicate, they would need to exchange their respective information. Moreover, only the time slots where both the nodes are free can be used for data transmission. This information also needs to be exchanged between the communicating nodes. Hence, while the measurement process does not involve any overheads, we still need to exchange some data among the wireless nodes.

References

[1] Dhruv Gupta, Daniel Wu, Prasant Mohapatra, Chen-Nee Chuah, “Experimental Comparison of bandwidth Estimation Tools for Wireless Mesh Networks”, in Infocom mini-conference, 2009

—

Passive bandwidth inference in wired networks

The research community has devoted a lot of effort to design novel bandwidth estimation tools during the last decade. As a result bandwidth measurements have become a mature research topic with well-developed results both at a fundamental and practical level. However the existing literature has shown that there is a clear trade-off between accuracy and intrusiveness. For instance PathLoad has to inject on average between 5MB and 10MB per measurement. Researchers have progressively reduced this huge amount of probe traffic and recent tools, such as IGI and Spruce, only need to transmit 130KB and 300KB, respectively. Although this load can be considered as negligible for the network, many applications (e.g. overlay routing) may continuously and simultaneously repeat AB measurements from a large set of end hosts. In addition, since these tools need to transmit probe traffic at very high rates, they can significantly impact the response time of existing connections. In particular, it has been observed that both tools (IGI and Spruce) can increase the response time of long TCP connections by a factor of 2-3.

At UPC we have developed a passive bandwidth inference methodology: PKBest (Passive Kalman-Based estimation). Our methodology is based on a Poisson model, but instead of injecting probe traffic, it uses the traffic already existing between two nodes to estimate the available bandwidth of the path between them. Therefore, its intrusiveness is negligible. Our methodology relies on the empirical observation that small sequences of Poisson-distributed traffic are naturally present in the Internet, which can be exploited for bandwidth estimation. The main challenge is that we cannot rely on any particular rate in the traffic, as in active tools. Instead, we estimate the linear equation rate versus utilization from these Poisson-distributed sequences.
As future work we plan to continue on research either low-intrusive or passive bandwidth estimation tools.