Methodology

We use ns 2.1b8a for our simulation experiments. The topology used is a bar-bell in which $N$ TCP senders transmit through a shared bottleneck link $L$ to an equal number of receivers. The router connecting the senders to $L$ becomes the bottleneck queue. Routers perform drop-tail FIFO queueing except in experiments with RED turned on. The buffer size is set to the bandwidth delay product. Packets are 512 bytes in size and the propagation delay is set to 50ms. We vary the capacity of the link in order to simulate different amounts of spare capacity.

We use a 15 minute section of a Squid proxy trace logged at UC Berkeley as the foreground traffic over L. The number of flows fluctuates as clients enter and leave the system as specified by the trace. On average there are about 12 active clients. In addition to this foreground load, we introduce permanently backlogged background flows. For the initial set of experiments we fix the bandwidth of the link to twice the average demand bandwidth of the trace. The primary metric we use to measure interference is the average transfer latency of a document i.e., the time between its first packet being sent and the receipt of the ack corresponding to the last packet. We use the total number of bytes transferred by the background flows as the measure of its utilization of spare capacity.

Unless otherwise specified, the values of the $threshold$ and $fraction$ for Nice are set to 0.1 and 0.5 respectively. We compare the performance of Nice to several other strategies for sending background flows. First, we compare with router prioritization that services a background packet only if there are no queued foreground packets. Router prioritization is the ideal strategy for background flow transmission, as background flows never interfere with foreground flows. In addition, we compare to Reno, Vegas( $\alpha=1, \beta=3$), Vegas( $\alpha=0, \beta=0$).