Flow-based bandwidth estimation in non-visible links

We address the problem of estimating the bandwidth of internal links in a network, based on flow records (e.g. NetFlow records) obtained in an external link, as described in Fig. 1. Performing such estimation has the advantage of requiring less measurement points and additional bandwidth, compared to a full deployment of NetFlow measurement points or a periodic polling of SNMP records on each interface of a router.

Fig. 1: Referential topology for the scope of this project. We aim to estimate bandwitdh of internal links ei based on measurements on external link e.

Our main challenge for this problem is to estimate the portion of traffic flowing among internal links ei, based on the portion of traffic going outside the newtork and captured at the measurement point. To achieve this goal, we are proposing a regression-based approach, which requires obtaining SNMP records of per-interface byte counts from the router directly connected to all links (router R1 in Fig. 1), but only during an initial training period. With these byte counts, we build a linear model that allows to estimate the full traffic on each link based on measurements performed at the external measurement point.

The tasks to be performed on this project include an experimental setup to validate the regression-based approach, and a study on the effect over the estimates' accuracy when using packet sampling to generate the flow records.

People

  • Prof. Mark Coates
  • Rodrigo Alarcon-Reyes (M. Eng. student)

 

References

  • Cisco Systems Inc., "Introduction to Cisco IOS NetFlow - A Technical Overview", last updated October 2007. Available online