Flux Research Group / School of Computing

Realistic Traffic Shaping in the Dummynet Link Emulator

Aisha Syed

Masters Thesis, University of Utah. August 2014.

Networking, Testbeds


Dummynet is a link emulator that can be used by itself, as well as integrated within testbeds such as Emulab. Despite its popularity in the research community, Dummynet still lacks the ability to precisely emulate certain real network effects. In particular, it has no support for packet reordering. Since reordering is a common and prevalent network phenomenon just like packet loss or delay, it cannot be ignored when implementing emulators if we want to provide realistic emulation.

It has been observed that networks suffer from reordering caused by packet striping, retransmissions, load balancing, multipath forwarding, etc. This has significant negative effects on the performance of both Transmission Control Protocol (TCP) and User Datagram Protocol (UDP). With the increase in prevalence of real-time streaming UDP applications such as video conferencing and Internet Protocol Television (IPTV), it has become important to focus on this problem which affects the performance of all these applications. Research into models and tools to diagnose and understand reordering requires that a sophisticated metric be used to describe it.

So, in this thesis, I make two contributions: improving the realism of traffic shaping in Dummynet emulator by adding support for emulation of reordering, and an algorithm, a max-flow solver, that generates reordered sequences to be used by Dummynet, from a sophisticated reordering metric called Reorder Density (RD). My implementation enables the user to specify the desired amount of reordering in a metric, such as RD (or even others), and have Dummynet generate traffic that is reordered according to the input metric’s value. This is accomplished within Dummynet by the use of a newly implemented scheduler.

I conclude my thesis with an evaluation using real and software generated network traces to show that the algorithm is scalable and the implementation works correctly. Also, a datapath evaluation to show that my modifications to Dummynet do not result in any unnecessary increase in emulation running time is included.