Empirical Analysis of TCP Losses and Its Detection/Recovery Mechanisms

While it is well-known that TCP performance degrades significantly on experiencing packet losses, not much is known about the way in which TCP losses occur in the real-world. In this paper, we address the questions: to what extent, and in what manner, do real-world TCP connections experience packet losses? And how effectively do TCP loss detection and recovery mechanisms deal with the packet losses? We answer these questions by conducting an extensive passive analysis of TCP packet traces. Our analysis attempts to (i) characterize the TCP loss process, (ii) evaluate the accuracy and timeliness of loss detection in TCP, and (iii) evaluate the timeliness and efficacy in avoiding further losses of the loss recovery mechanisms in TCP. To facilitate this analysis, we first implement detailed sender-side state machines for several prominent TCP stacks (Windows, Linux, BSD, and Solaris) and augment these with extra logic to correctly track TCP sender state as well as actual packet losses. Using these state machines we analyze traces of more than TCP connections, collected from different locations around the world. Our analysis sheds several insights on TCP loss detection and recovery. of TCP loss detection is triggered due to timeouts. 90% of detection using duplicate acks take place within 2 RTTs. TCP experiences no more than consecutive losses of the time; however, 20-30% of lossrecovery periods last more than RTTs. Our results, thus, confirm the efficiency of some TCP mechanisms and highlight the deficiency of others. This research was supported in part by NSF CAREER grant CNS0347814, a UNC Junior Faculty Development Award, and NSF RI grant EIA-0303590.