We present the concept of network traffic streams, and the ways they aggregate into flows through Internet links. We describe a method of measuring the size and lifetime of Internet streams, and use this method to characterise traffic distributions at two different sites.

As the most widely used reliable transport in today’s Internet, TCP has been extensively studied in the past decade. However, previous research usually only considers a small or medium number of concurrent TCP connections. The TCP behavior under many competing TCP flows has not been sufficiently explored. In this paper 1, we use extensive simulations to systematically investigate the performance of a large number of concurrent TCP flows. We start with a simple scenario, in which all the connections have the same round-trip time, and the gateways use Drop-Tail policy. We examine how the aggregate throughput, goodput, and loss rate vary with different underlying topologies. We also look at the behavior of each individual connection when competing with other connections. We observe global synchronization in some cases. We break the synchronization by either adding random processing time or using RED gateways, and examine their effects on the TCP performance. Finally we investigate the TCP performance with different RTT’s, and quantify the roundtrip bias using both analysis and simulations. Keywords: TCP, congestion control, drop-tail, RED, simulation.

Several gateway-based congestion control mechanisms have been proposed to support an end-to-end congestion control mechanism of TCP (Transmission Control Protocol) . One of promising gateway-based congestion control mechanisms is a RED (Random Early Detection) gateway. Although effectiveness of the RED gateway is fully dependent on a choice of control parameters, it has not been fully investigated how to configure its control parameters. In this paper, we analyze the steady state behavior of the RED gateway by explicitly modeling the congestion control mechanism of TCP. We first derive the equilibrium values of the TCP window size and the buffer occupancy of the RED gateway. Also derived are the stability condition and the transient performance index of the network by using a control theoretic approach. Numerical examples as well as simulation results are presented to clearly show relations between control parameters and the steady state behavior. Our findings are: (1) max p (maximum packet marking probability) mostly affects the RED's buffer occupancy, (2) the network becomes more stable as the number of TCP connections or the bandwidth--delay product increases, and (3) min th (minimum threshold) is a key parameter for optimizing the transient performance. We finally discuss how control parameters of the RED gateway should be configured for achieving better performance.

Previous studies of Internet traffic have shown that a very small percentage of flows consume most of the network bandwidth. It is important to understand the characteristics of such flows for traffic monitoring and modeling purposes. Several prior researchers have characterized such flows using different classification schemes: by size as elephant and mice; by duration as tortoise and dragonfly; and by burstiness as alpha and beta traffic. However, it is not clear how these different definitions of flows are related to each other. In this work, using data recorded from two different operational networks, we study these “heavy-hitter ” flows in four different dimensions, namely size, duration, rate and burstiness, and examine how they are correlated. This paper makes three contributions: First, we systematically characterize prior definitions for the properties of such heavy-hitter traffic. Second, based on our datasets, we observe that there are strong correlations between some combinations of size, rate and burstiness. Finally, we provide a plausible explanation for the observed correlations. We show that these correlations could be explained by transport and application-level protocol mechanisms.

We present and discuss stream size and lifetime distributions for web and non-web TCP traffic on a campus OC12 link at UC San Diego. The distributions are stable over long periods, and show that on this link only 3% of the streams last longer than one minute, and that only about 0.5% of them are bigger than 100 kBytes. Although there are large streams (elephants) on this link, the bulk of its traffic is composed of many small streams (mice).

In this survey, we first review the concept of congestion control with a focus on the Transmission Control Protocol/Internet Protocol (TCP/IP). We describe many recently proposed algorithms to combat congestion and improve performance, particularly active queue management (AQM) algorithms such as random early detection (RED) and its variants. We then survey control-theoretic analysis and design of TCP congestion control with an AQM scheme. In addition, we discuss three problems associated with AQM proposals: parameter setting, the insensitivity to the input traffic load variation, and the mismatch between macroscopic and microscopic behavior of queue length dynamics. As alternatives to AQM algorithms, we also survey architectural approaches such as modification of source or network algorithms, and economic approaches including pricing or optimization of allocated resources. Finally, we list many open issues that persist in the design, operation, and control of the Internet. Internet congestion occurs when the aggregate demand for

Abstract not found

We assess the state of Internet congestion control and error recovery through a controlled study that considers the integration of standards-track TCP error recovery and both TCP and router-based congestion control. The goal is to examine and quantify the benefits of deploying standards-track technologies for contemporary models of Internet traffic as a function of the level of offered network load. We limit our study to the dominant and most stressful class of Internet traffic: bursty HTTP flows. Contrary to expectations and published prior work, we find that for HTTP flows (1) there is no clear benefit in using TCP SACK over TCP Reno, (2) unless congestion is a serious concern (i.e., unless average link utilization is approximately 80 % or higher), there is little benefit to using RED queue management, (3) above 80 % link utilization there is potential benefit to using Adaptive RED with ECN marking, however, complex performance trade-offs exist and the results are dependent on parameter settings. * I.

The Internet is a heterogeneous environment comprising wired and wireless components, and transport protocols with different error recovery strategies. We use simulations to study the behaviour of random early detection (RED) gateways in such heterogeneous environments. We investigate two issues: (i) the performance trade-offs of the dropping policy of RED over Drop-Tail’s, and (ii) the impact of RED’s active queue management on more sophisticated protocols such as TCP-SACK. Some of our results indicate that RED does not deal adequately with heterogeneity. In particular, it may happen that RED applies congestion avoidance techniques even when wireless errors force some senders to back off prior to actual congestion. Our scenarios involve a non-monotonic transmission behaviour of transport protocols within one and the same congestion epoch; RED may escalate further bandwidth under-utilization by applying false congestion avoidance tactics. Furthermore, our results indicate that RED’s dropping policy of ‘proportional ’ fairness, which is realized through a Send-more/Drop-more scheme, can also penalize sophisticated protocols like TCP-SACK, which attempt to recover more aggressively from wireless losses. In summary, our results call for further investigation of RED’s efficiency in heterogeneous networks and naturally raise the concern of RED’s compliance with the end-to-end argument. Copyright # 2004 John

Network models today are often derived from two di#erent methods. On one hand, detailed tra#c models are generated based on traces from a single tap into the network. Alternatively, one can collect higher-level tra#c-matrix data with SNMP from many routers. However, inferring flow-level details from such data is still an open research issue. Today it is infeasible to collect a fine-grained, packet-level representation of a complete, multi-router network. Even if it were economically feasible to synchronize and monitor every router in a large network, the amount of data generated would tax storage and computation resources. In this work, we propose a methodology to infer flow-level tra#c across a network by exploiting the correlations between user populations across di#erent networks. The contribution of this paper is threefold. First, we provide a formal description by which one can quantify the level of "similarity" between two networks. Second, we show that the user-behavior parameters of the tra#c (such as user "think" time in web tra#c) are correlated across time, while the application-specific parameters of the tra#c (such as object size) are correlated across "similar" networks. Finally, by utilizing the correlations between similar networks, we present a methodology for inferring tra#c at places where continuously taking measurements are infeasible. We then demonstrate our approach via a case study on traces of web tra#c and evaluate its e#ectiveness via analysis and simulation.