Detecting Shared Bottlenecks

Sridhar Machiraju
(Professor Randy H. Katz)

It is known that alternative paths to normal Internet routes could be used to improve connectivity. This can be achieved through well placed overlay routers. In order to prevent interference with each other, different overlay flows would benefit from knowledge of bottlenecks that they share. We are currently investigating techniques by which we can decide if two flows share the same bottleneck. Our goal is to do make this decision based on the correlation of the throughputs observed by the flows. Detecting such correlation is complicated because the short-term dependence of TCP throughput on packet drops is not very well understood. Also, the packet drops themselves may not be evenly distributed among flows in case of droptail queues.


Send mail to the author : (machi@eecs.berkeley.edu)


Edit this abstract