Rescheduling Lengthy Downloads on Bottleneck Links

Abstract

TCP’s ability to share bottleneck bandwidth fairly and efficiently dwindles when faced with an increase in the number of competing flows. Indeed studies have shown that long TCP flows consume more than 90% of the available bandwidth, with short flows consuming a very small portion. It is also known that out of these, a very small percentage of the largest flows carry the majority of the bytes. The utilization pattern of bottleneck links exacerbates this situation. During the daytime (peak period), the network utilization is characteristically high whereas at night (off-peak period), the link is virtually idle. At peak periods, TCP’s congestion control mechanisms highly favour long-lived flows at the expense of their short-lived counterparts. We discuss the design and implementation of a flow rescheduling tool. The tool identifies, blocks and reschedules long flows to reconnect during the off-peak period. Network bottlenecks are a commonplace phenomenon in many IP networks just as much as they are unavoidable. TCP’s ability to share bottleneck bandwidth fairly and efficiently dwindles when faced with an increase in the number of competing flows. Indeed recent studies have shown that long TCP flows consume more than 90% of the available bandwidth, with short flows consuming a very small portion. It has also been realized that out of these, a very small percentage of the largest flows carry the majority of the bytes. The utilization pattern of bottleneck links does not make matters any better. During the day, the utilization pattern is characteristically high whereas at night, the link is virtually idle. We refer to these times as the peak and off-peak periods respectively. At peak time, TCP’s congestion control mechanisms highly favor long-lived flows at the expense of their short-lived counterparts. This work discusses the design and implementation of a flow rescheduling tool, specifically illustrating the techniques used to achieve accuracy and robustness. The tool works by identifying and rescheduling long flows to reconnect during the off-peak period, when link utilization is very low. We present validation results for the tool which demonstrate its reliability in the face of actual Internet conditions.