Bandwidth estimation: metrics, measurement techniques, and tools

Mon, 03/17/2008 - 11:04 by Damien Saucez • Categories:

The paper titeled "Bandwidth estimation: metrics, measurement techniques, and tools" (R Prasad, C Dovrolis, M Murray, K Claffy - Network, IEEE, 2003) is divided in three main chapters. First, it defines the three following metrics:
- Capacity: the maximum possible IP layer transfer rate.
- Available bandwidth: the unused, or spare, capacity of a link during a certain time period.
- Bulk Transfer Capacity (BTC): the maximum throuhput obtainable by a single TCP connection (depends on many factor like rtt, window size, TCP implementation).

Second, a brief overview of the current techniques to measure these metrics is presented.

The more important techniques are VPS (variable Packet Size) which estimates the capacity of individual hops. On the contrary PPTD (Packet Pair/Train Dispersion) estimates the end-to-end capacity while TOPP (Trains Of Packet Pairs estimates the end-to-end available bandwidth. Unfortunataly, no technique is presented (known?) to estimate available bandwidth of individual hops.

VPS measure the RTT and infer the capacity by considering the minimum RTT of a packet of size L is: $RTT(L) = \alpha + \beta L$ where $\beta$ is a function of the capacity described in detail in the paper.

The PPTD uses the packet dispertion to estimate the available bandwith: the dispertion is a function of the capacity. To estimate the dispertion, the technique sends multiple pairs of back to back constant size packets. The packet pair dispertion gives the dispertion.

The Self-Loading Periodic Streams uses one-way delay to estimate the available bandwidth. The concept is simple, just send data at a given rate and measure the one-way delay. If the delay increases, it means that the rate is higher than the rate supported.

TOPP sends trains of packets with gradually increasing rates. Base on the dispertion at the receiver, it is possible to infer the available bandwidth. When the rate is more important than the supported rate, queuing occures and the train's dispertion is affected.

The end of the paper presents a taxonomy of the current bandwidth measurement tools.

The paper is a good introduction to bandwidth estimation with clear descriptions of the concepts introduced. Unfortunately, the paper only introduces the problem but good references gives the possibility to study the different concepts in detail.