Torchestra: reducing interactive traffic delays over tor

Tor is an onion routing network that protects users' privacy by relaying traffic through a series of nodes that run Tor software. As a consequence of the anonymity that it provides, Tor is used for many purposes. According to several measurement studies, a small fraction of users using Tor for bulk downloads account for the majority of traffic on the Tor network. These bulk downloads cause delays for interactive traffic, as many different circuits share bandwidth across each pair of nodes. The resulting delays discourage people from using Tor for normal web activity.

We propose a potential solution to this problem: separate interactive and bulk traffic onto two different TCP connections between each pair of nodes. Previous proposals to improve Tor's performance for interactive traffic have focused on prioritizing traffic from less active circuits; however, these prioritization approaches are limited in the benefit they can provide, as they can only affect delays due to traffic processing in Tor itself. Our approach provides a simple way to reduce delays due to additional factors external to Tor, such as the effects of TCP congestion control and queuing of interactive traffic behind bulk traffic in buffers. We evaluate our proposal by simulating traffic using several methods and show that Torchestra provides up to 32% reduction in delays for interactive traffic compared to the Tor traffic prioritization scheme of Tang and Goldberg [18] and up to 40% decrease in delays when compared to vanilla Tor.