The Local and Global Effects of Traffic Shaping ( Extended Version )

The Internet is witnessing explosive growth in traffic, in large part due to bulk transfers. Delivering such traffic is expensive for ISPs because they pay other ISPs based on peak utilization. To limit costs, many ISPs are deploying ad-hoc traffic shaping policies that specifically target bulk flows. However, there is relatively little understanding today about the effectiveness of different shaping policies at reducing peak loads and what impact these policies have on the performance of bulk transfers. In this paper, we compare several traffic shaping policies with respect to (1) the achieved reduction in peak network traffic and (2) the resulting performance loss for bulk transfers. We identified a practical policy that achieves peak traffic reductions of up to 50% with only limited performance loss for bulk transfers. However, we found that the same policy leads to large performance losses for bulk transfers when deployed by multiple ISPs along a networking path. Our analysis revealed that this is caused by certain TCP characteristics and differences in local peak utilization times. We present a data staging service that counteracts the deleterious end-to-end effects of local traffic shaping by delaying bulk data at appropriate points in the network to consume bandwidth when available. We show that the service is incrementally deployable and has reasonable storage requirements.