On the impact of TCP and per-flow scheduling on Internet performance (extended version)

TL;DR

This paper models the effects of per-flow scheduling and buffer management on TCP/UDP performance using fluid dynamics, providing analytical insights into throughput, fairness, and loss rates in Internet traffic.

Contribution

It introduces a novel fluid model for analyzing TCP and UDP flows under per-flow scheduling, deriving closed-form expressions for key performance metrics.

Findings

Analytical expressions for TCP throughput and buffer occupancy.

Insights into fairness between TCP flows under different scheduling.

Characterization of UDP loss rates in mixed traffic environments.

Abstract

Internet performance is tightly related to the properties of TCP and UDP protocols, jointly responsible for the delivery of the great majority of Internet traffic. It is well understood how these protocols behave under FIFO queuing and what the network congestion effects. However, no comprehensive analysis is available when flow-aware mechanisms such as per-flow scheduling and dropping policies are deployed. Previous simulation and experimental results leave a number of unanswered questions. In the paper, we tackle this issue by modeling via a set of fluid non-linear ODEs the instantaneous throughput and the buffer occupancy of N long-lived TCP sources under three per-flow scheduling disciplines (Fair Queuing, Longest Queue First, Shortest Queue First) and with longest queue drop buffer management. We study the system evolution and analytically characterize the stationary regime: closed-form expressions are derived for the stationary throughput/sending rate and buffer occupancy which give thorough understanding of short/long-term fairness for TCP traffic. Similarly, we provide the characterization of the loss rate experienced by UDP flows in presence of TCP traffic.