Although a significant amount of effort has been directed at discovering attacks against anonymity communication networks and developing countermeasures to those attacks, there is little systematic analysis of the Quality of Service (QoS) for such privacy preserving systems. In this paper, we initiate the effort to study the QoS of Tor, a popular anonymous communication network on the Internet. We find that Tor suffers severe TCP performance degradation because of its random path selection strategy. Our investigation shows that Tor’s bandwidth weighted path selection algorithm can only improve the performance to a very limited extent. We analyze this performance issue from the perspective of overlay networks and model the TCP throughput of Tor. We conduct extensive experiments on the real-world Tor network and the experimental results validate our theory. We also discuss possible remedies to this performance issue.