How loss and delay correlations affect TCP pseudo self-similarity

In this paper we exploit a model of a single TCP source introduced in [1] to show the dependence of the scaling properties of the TCP throughput on the correlation structure of the Round Trip Times (RTTs) and of the packet losses. The source works according to the most common TCP mechanisms, i.e. slow-start, congestion avoidance and exponential back-off retransmission timer. The innovation of the model consists in the representation of the network feedback on the source: we use suitable list of RTTs to feed the model simulating the TCP source, as the statistics of the RTTs reflect the network collective behavior, including the congestion periods. Using this model we have carefully analyzed the second-order scaling behaviour of the throughput of the single TCP source in a limited range of time scales, i.e. its pseudo self-similarity. We show that the pseudo self-similarity of the throughput depends significantly on the correlation structure of the losses. In order to support our findings, we have considered lists of RTTs measured in real networks as well as synthetic traces with suitable correlation structures.