Making a Delay-based Protocol Adaptive to Heterogeneous Environments Kiran Kotla and A. L. Narasimha Reddy Texas A&M University {kiran@cs, reddy@ece}.tamu.edu Abstract— This paper investigates the issues in making a delay-based protocol adaptive to heterogeneous environments. We address how a delay-based protocol can compete with a loss-based protocol such as TCP. We investigate if potential noise and variability in delay measurements in environments such as cable and ADSL access networks impact the protocol behavior significantly. We investigate these issues in the context of incremental deployment of a new delay-based protocol, PERT. We propose design modifications to PERT to compete with SACK. We show that PERT experiences lower drop rates than SACK and leads to lower overall drop rates with different mixes of PERT and SACK protocols. Second, we show that a single PERT flow can fully utilize a high-speed, high-delay link. The results from ns-2 simulations indicate that PERT can adapt to heterogeneous networks and can operate well in an environment of heterogeneous protocols. We also show that proposed changes retain the desirable properties of PERT such as low loss rates and fairness, when operating alone. The protocol has also been implemented in the Linux kernel and tested through experiments on live networks, by measuring the throughput and losses between nodes in our lab at TAMU and different machines on the planet-lab. Keywords-Congestion protocol, delay-based, ns-2 simulations, emulations, incremental deployment, high-speed links I. INTRODUCTION Multimedia applications have low-delay and low-loss requirements. Currently, applications either use UDP or TCP as transport protocols. UDP does not provide congestion control and applications employing UDP have to incorporate congestion control. TCP’s in-order reliable delivery may not be necessary for multimedia applications. A number of congestion control approaches have been proposed to replace TCP to move networks towards higher multimedia content and higher link bandwidths in the future. Traditionally, congestion protocols have taken two different approaches of inferring congestion from network feedback – (a) through packet losses or markings at the router, or (b) from measured characteristics such as delays at end hosts. Challenges in estimating delays accurately and other issues have resulted in skepticism of viability of delay-based schemes [1], [2]. Recently, we addressed some of these issues and proposed a delay-based protocol, PERT (Probabilistic Early Response TCP) [3]. PERT improves the delay estimation process and deals with remaining uncertainties through a probabilistic response to measured delays. PERT emulates AQM's behavior at end hosts, responding at a higher rate at higher delays. While PERT has been shown to be effective in reaching its goals, a number of technical challenges remain in its practical deployment. First and foremost is the issue of how delay-based protocols can compete with various versions of TCP. Delay-based protocols, by responding to congestion early, cede ground to loss-based protocols that keep increasing their rate until a packet is dropped (as most versions of TCP do). While most delay-based protocols exhibit good properties in a homogenous deployment, for the delay-based protocols to be practical, they need to be able to operate in an environment of mixed protocol deployment. This would be necessary for incremental deployment. We address this issue (and others) in this paper. We propose and evaluate design enhancements to enable the delay-based protocol, PERT to compete with TCP-SACK. We also study what advantages and benefits may be realized as a mix of PERT and SACK flows evolves from 100% SACK to 100% PERT. We show through these experiments that incremental deployment of PERT provides lower overall drop rates along with lower drop rates for the PERT flows. While some recent protocols [5, 18, 21] have strived for coexistence with TCP, we are not aware of any work that simultaneously deals with incremental deployability of a new protocol and fair bandwidth sharing with TCP in mixed protocol deployment workloads. A second issue that has been raised in the past regarding delay-based protocols is their robustness to noise in delay measurements. This has been the primary motivation for employing the probabilistic response in PERT [3]. Recent work on cable and ADSL access networks has highlighted the RTT variance of these networks even in the absence of congestion [17], due to their network access granting and scheduling mechanisms. This raises the question whether delay-based congestion protocols can function effectively when deployed in such access networks. We study this issue through practical deployment of a delay-based protocol in cable and ADSL networks. We report on PERT’s ability to correctly gauge congestion in networks with widely varying access delays. We also evaluate PERT's robustness to measurement noise by deliberately adding noise to measured delays in simulations. These evaluations show that PERT can be more robust to noise in delay compared to FAST and Vegas. A third issue we address is whether a delay-based protocol can be scaled to provide high utilizations for single flows in high-speed, high-delay links. This has been a topic of considerable interest lately and many new protocols have been proposed [4, 5, 6, 7, 8]. We show that PERT can fully utilizehigh-speed, high-delay network links and provide much lower drop rates than loss-based schemes such as [4, 6, 7]. While FAST [5], a delay-based protocol, has been designed to compete with TCP and operate in high-speed networks, not much work has been reported on its performance in environments of mixed protocol deployment. Our work here emphasizes this aspect. We compare PERT’s performance with that of FAST in incremental deployment scenarios. We propose, analyze and evaluate design modifications for PERT to deal with these important issues. The suggested modifications are simple to implement and are shown to be effective through analysis, ns-2 based simulations and testing our Linux based kernel implementation over the Internet. The paper makes the following significant contributions: (a) Adapted a delay based protocol to be competitive with the existing loss based congestion protocol TCP, (b) conducted extensive evaluations of mixed deployment scenarios of