A root cause analysis toolkit for TCP
Computer Networks: The International Journal of Computer and Telecommunications Networking
TCP revisited: a fresh look at TCP in the wild
Proceedings of the 9th ACM SIGCOMM conference on Internet measurement conference
On the interaction between internet applications and TCP
ITC20'07 Proceedings of the 20th international teletraffic conference on Managing traffic performance in converged networks
| Hi-index | 0.00 |
We propose a methodology to infer end-end path properties of TCP connections (such as packet loss, reordering and delay) and other factors that affect TCP sender behavior using passive measurements collected at a single point of observation along the end-end path. With this passive approach, an observation point in a Tier-1 backbone network can observe and analyze millions of TCP connections, which originate and terminate from a highly diverse cross-section of end points in today's Internet - a capability that is unmatched by current active measurement techniques. We apply our passive measurement inference techniques on traces collected in a large Tier-1 backbone network. We analyze the causes of out-of-sequence packets (due to loss, reordering and in-network duplication), the distribution and variation of round-trip delay, the congestion control flavors of TCP, and the extent to which application-level considerations limit TCP throughput. We validate the accuracy of our measurement-based inference techniques by comparing inferred behavior with active measurements made on monitored flows. But empirical validation alone is insufficient, since monitored paths can exhibit a wide range of network properties, many of which may occur only rarely. Thus, we use a combination of model-checking and formal reasoning to identify all possible events in the network for which the inference rules can produce incorrect results.