CPSC538A Paper presentation 2
Passive Inference of Path Correlation
Paper Summary:
This paper presented the two algorithms based on passive measurement of TCP throughput to determine the level of shared congestion. It does not send any probe traffic as many active measurements do, it only passively gathers the information, gradually learning of congested links and sharing that information with other nodes in the system.
For the Capacity Conservation algorithm, they reduce the rate of one of the flows by a known amount for a fixed time period (they turn the flow off) and then later returns the flow to its normal rate( they turn the flow back on). During this period, they observe whether the other flow change constantly and high percentage of slopes that are negative. If so, they consider the two flows are correlated. It works well when the background traffic does not change frequently and the total number of background flows is small.
In order to solve that limitation, they introduced the Correlated Variation algorithm. It changes the background traffic randomly, if the foreground flow also changed randomly, they say they are sharing the same bottleneck. They test the algorithm both on Emulab and Planetlab to verify the approach.
The two algorithms are complementary to each other, and all the algorithms are based on busy source overlay networks.
Discussion Summary:
In algorithm 1, it seems not practical to turn on and off the real video flow for observing the correlation in real world.
For algorithm2, how to randomly set the real stream flow into randomly change is a question. And also how much they need to exchange the nodes to get the best correlation coefficient is also a vague problem.
Although they used the Lazy-evaluation approach to decrease the computational complexity from O(mN