160 likes | 266 Views
Is the Round-trip Time Correlated with the Number of Packets in Flight?. Saad Biaz, Auburn University Nitin H. Vaidya, University of Illinois at Urbana-Champaign. published in ACM IMC ’ 03. Overview. TCP Reno keeps increasing sending rate till packet loss detected
E N D
Is the Round-trip Time Correlated with the Number of Packets in Flight? Saad Biaz, Auburn University Nitin H. Vaidya, University of Illinois at Urbana-Champaign published in ACM IMC’03
Overview • TCP Reno keeps increasing sending rate till packet loss detected • some Congestion Avoidance Techniques (CATs) are proposed to avoid such aggressive behavior • some CATs use RTT / throughput to detect queue build-up and adjust congestion window size • Can RTT reflect queue build-up?
Outline • Terminology • Correlation coefficients used • Experimental data • Analysis results • Conclusion
Terminology Pi , PWi , PFi , PTi , W(Pi) / Wi , RTT(Pi) / RTTi , Sign(X) , Cl , B(Cl)
Correlation Coefficients • Similar correlation coefficients for direction change (increase or decrease)
TCP-Vegas • compare the quantity with two thresholds, alpha and beta (alpha < beta) • if the quantity < alpha, cwnd increases • if the quantity > beta, cwnd decreases • otherwise, cwnd holds constant
Experiments • Data is collected by Paxson in 1994-95 • span many paths with different RTT • two data sets N1 and N2 are collected, which are tcpdump traces over 37 sites across USA, Europe and Australia • this study uses only N2 of 20000 tcpdump traces for bulk transfers of 100 KB between pairs of sites among the 37 concerned sites • N2 contains both the tcpdump traces at sender and receiver, but only sender is consider in this study
Experiments • 14218 out of 20000 tcpdump traces are analyzed for technical reasons • These 14218 connections were between 31 different sites • Not all sites send to all sites; span only 737 paths • RTT and data in flight (W) are found from those tcpdump traces
Partitioning connection sets • to see if and how bottleneck bandwidth affects correlation coefficients, the TCP connections are divided into two sets: fast and slow
Correlation bwt RTT and W Slow Fast frequency distribution
From corrcoef in direction of change in RTT and W for PW, about 88% connections have positive correlation In general, RTT is sensitive to load However, corrcoef is not strong enough to be relied (on RTT) to build congestion avoidance scheme 88% Conclusion
Conclusion • RTT observed by a TCP connection is dependent on other traffic; effect is more significant if that TCP connection only share a small fraction of bandwidth • vagaries of transport protocol like delayed ACK