Machine Learning-Based Packet Dropping Schemes
Kurtis Heimerl and Randy H. Katz
Router congestion is still a problem for large-scale Internet services. Congestion has taken down systems as large as Amazon.com and Skype. In this research I aim to utilize machine learning to determine the relative importance of packets. With this priority information we can make more informed decisions regarding packet dropping during congestion events.
To make the packet priority determinations, I propose an active testing methodology. We will use the Click modular router  to implement a testing framework. We will destroy existing TCP connections that pass through the router and make determinations based on the packet flow that follows. If we determine a system has died due to the loss of the connection, we then add a dependency to our graph. These system failures are reasonable due to the reliability framework of the data center.
Once the dependency graph is complete, we can use it to maximize the number of leaves in the graph. This gives us the largest number of services running if the network becomes congested, and a level of stability that current large scale distributed systems are lacking.
- E. Kohler, R. Morris, B. Chen, J. Jannotti, and M. F. Kaashoek, ACM Transactions on Computer Systems, Vol. 18, No. 3, August 2000, pp. 263-297.