Theoretical and experimental evaluation of the two-level processor sharing discipline for TCP flows

Size-based scheduling policies have been widely studied in the literature, and their interest in networking applications has been huge in the last decade. These policies consist in deciding the priority of the packets belonging to a certain flow based on the service time that the flow has received up to a certain epoch or, when possible, on the remaining service time. The scientific literature has devoted many efforts to the comparison and analysis of these disciplines either by simulation or by analytical models although real-world implementations can have different performance for numerous reasons. In this paper, we consider the two-level processor sharing discipline (2LPS), i.e., packets are served with high priority if they belong to a flow that has required less than a work up to that moment, where a is a threshold-parameter of the model. The goal is that of assessing the performance of this discipline once implemented in a real router with a real-world network traffic and compare these measurements with the performance indices obtained by the queueing model. To characterise the networks traffic, we fit two datasets with an acyclic phase-type distribution thanks to an existing tool and then transform the resulting distribution into a generalised hypergeometric distribution. Our experiments confirm that the 2LPS improves the flow expected response time with respect to the standard scheduling by taking advantage of the heavily tailed distribution characterising the TCP flow sizes, but this improvement seems slightly smaller than what predicted by the analytical models