Does TCP Produce Self-Similar Traffic?

Don Towsley

Dept. of Computer Science, Univ of Massachusetts


In this talk we examine the correlation structure of traffic generated by a single TCP session. Contrary to recent claims made in the literature that TCP congestion control creates self-similar traffic, we show here that under realistic conditions, the traffic generated by a long-lived TCP connection, while exhibiting pronounced correlations over a finite range of time sales, is not consistent with long-range dependence or (asymptotic) self-similarity. Our conclusion is supported by a careful statistical analysis of TCP trace data and by analytical results derived from a versatile and well-established Markovian model of TCP. In particular, we analyze the same TCP traces used in an earlier study to ``prove'' that TCP creates self-similar traffic. By avoiding a number of pitfalls associated with inferring self-similar characteristics, we arrive at a conclusion that directly contradicts the claim of self-similarity but that fully conforms to the behavior predicted by the analytic TCP model. This work points out the importance of careful trace analysis and detailed examination (and cross validation) of alternative explanations when establishing or characterizing the generality of any particular finding about Internet traffic.

Joint work with A. Feldman, R. Figueiredo, B. Liu, W. Willinger

[Don Towsley]
[Dept. of Computer Science, Univ of Massachusetts]