Abstract In this paper we investigate the network performance of TCP hosts and gateways employing New Reno TCP using a parallel simulator. The gateway adopts the random early detection (RED) scheme and proactive congestion control mechanism to enforce fair sharing among competing TCP connections. The test network is modeled using PARSEC. The simulation reveals that the RED gateway drops only a small number of packets, and utilization of the bottleneck link on the RED gateway is less than 75%. The number of retransmission timeouts recorded at the TCP senders is also small. Moreover, increasing the bandwidth of the gateway input links does not substantially improve the network performance, while increasing the number of input links is very effective. Key words: congestion control, New Reno TCP, parallel simulation, PARSEC, random early detection. 1. Introduction The Internet has experienced explosive growth in the past few years. In the presence of more hosts and users with fast link connections, the traffic load has shown dramatic increase. Therefore Internet users often encounter some congestion. The need for Internet congestion control became apparent during mid-80’s, and network collapse due to congestion had already been predicted [1]. The random early detection (RED) scheme [2,3] was introduced as a measure of proactive congestion control for achieving low delay and high throughput. Study on the network performance of the TCP hosts and gateways employing the RED approach (shortly RED gateways) using a parallel simulator is the main objective of this paper. A significant amount of today's Internet traffic including WWW (HTTP), file transfer, email, and remote access, are carried by the TCP [4]. Therefore, traffic dynamics in the Internet are heavily influenced by the behavior of the TCP. Here the congestion control mechanisms are implemented at the hosts in the form of flow control. In 1988, Jacobson [5] pioneered the study of the TCP congestion control mechanisms; slow-start, congestion avoidance, conservation of packets, and exponential timer backoff. The TCP was later augmented with the fast retransmit and fast recovery algorithm to improve the efficiency [6,7,8]. However, there still exists a limit for the end-systems in controlling congestion. RED gateway was thus introduced to complement the end-system congestion control mechanisms [2,9]. The PARSEC (PARallel Simulation Environment for Complex system) [10] is a C-based language developed for parallel discrete event simulation based on the process-interaction approach. Here an object is represented by a logical process. Interactions among the processes are modeled by time-stamped message exchanges among the corresponding logical processes. It offers good support for high-level (application level) simulations with message passing infrastructure and abstraction. Therefore it is an efficient tool for simulating a system consisting of a number of concurrent processes such as network and operating system. In this paper we study the dynamics between the TCP hosts and RED gateway using PARSEC. The simulation results reveal that the RED gateway drops only a small number of packets for reasonable traffic condition and network structure. Utilization of the bottleneck link is less than 75%, and the number of retransmission timeouts recorded at the TCP sender is small. It was also identified that increasing the bandwidth of the gateway input links does not improve the network performance, while increasing the number of input links is very effective. The rest of the paper is organized as follows. Section 2 gives a brief introduction to the TCP with an emphasis on congestion control and the UCLA PARSEC language. Section 3 describes the modeling of the test network using PARSEC. Section 4 presents the simulation results, and Section 5 concludes the paper. Evaluation of Proactive Congestion Control on RED Gateway Kuoling Fang Alcatel USA, Inc. Plano, TX [email protected] Hee Yong Youn and Hyunseung Choo School of Electrical and Computer Engineering Sungkyunkwan Univ., Suwon, Korea {youn,choo}@ece.skku.ac.kr Chansu Yu School of Engineering Info & Comm Univ., Taejon, Korea [email protected]. TCP Congestion Control and PARSEC 2.1 New Reno TCP The TCP was designed to operate reliably over almost any transmission medium regardless of the transmission rate, delay, corruption, etc. The current TCP implementations adapt to the transfer rates in the range of 100 to 107 bps and round-trip delays of 1 ms to 100 seconds. The studies on the TCP performance have shown that it can work well over a variety of Internet paths, ranging from 300 bit/sec dial-up modems to 800 Mbit/sec I/O channels [11]. There are four intertwined congestion control algorithms in the TCP; slow start, congestion avoidance, fast retransmit, and fast recovery. The underlying mechanisms for these congestion control methods are congestion window and acknowledgement (ACK) [12]. Different versions of the TCP vary with the employed congestion control algorithms [4]. The version of TCP adopted in our simulation is New Reno TCP [13,14]. The features for congestion control employed in it are outlined below. • Slow-start and congestion avoidance. • Fast retransmission with the congestion window reduced to half and the slow-start threshold set to the new congestion window size. • Fast recovery with the inflation of the congestion window by the number of duplicated ACKs until a new ACK arrives. Exit of fast recovery only upon arrival of the ACK for the highest number segment transmitted. • Response to a partial ACK inferring that the segment has been lost. • Delayed acknowledgment as one ACK for every two segments received. 2.2 RED Gateway The traditional approach for managing the gateway queue length is to accept packets until the preset maximum length is reached, and then drop subsequent incoming packets until the queue length decreases. There are two important drawbacks in this approach. First, it allows a single connection or a few connections to monopolize