Improving explicit congestion notification with the mark-front strategy

Abstract

Delivering congestion signals is essential to the performance of networks. Current TCP/IP networks use packet losses to signal congestion. Packet loss not only reduces TCP performance, but also adds large delay. Explicit congestion notification (ECN) delivers a faster indication of congestion and has better performance. However, current ECN implementations mark the packet from the tail of the queue. In this paper, we propose the mark-front strategy to send an even faster congestion signal. We show that mark-front strategy reduces buffer size requirement, improves link efficiency and provides better fairness among users. Simulation results that verify our analysis are also presented.

Introduction

Delivering congestion signals is essential to the performance of computer networks. In TCP/IP, congestion signals from the network are used by the source to determine the load. When a packet is acknowledged, the source increases its window size. When a congestion signal is received, its window size is reduced [1], [2].

TCP/IP uses two methods to deliver congestion signals. The first method is timeout. When the source sends a packet, it starts a retransmission timer. If it does not receive an acknowledgment within a certain time, it assumes that congestion has happened in the network and the packet has been lost. Timeout is the slowest congestion signal because the source has to wait a long time for the retransmission timer to expire.

The second method is loss detection. In this method, the receiver sends a duplicate ACK immediately on reception of each out-of-sequence packet. The source interprets the reception of three duplicate acknowledgments as a congestion packet loss. Loss detection can avoid the long wait of timeout.

Both timeout and loss detection use packet losses as congestion signals. Packet losses not only increase the traffic in the network but also add large transfer delay. The explicit congestion notification (ECN) proposed in [3], [4] provides a light-weight mechanism for routers to send a direct indication of congestion to the source. It makes use of two experimental bits in the IP header and two in the TCP header. When the average queue length exceeds a threshold, the incoming packet is marked as congestion experienced, with a probability calculated from the average queue length. When the marked packet is received, the receiver marks the acknowledgment using an ECN-Echo bit in the TCP header to send congestion notification back to the source. Upon receiving the ECN-Echo, the source halves its congestion window to help alleviate the congestion.

Many authors have pointed out that marking provides more information about the congestion state than packet dropping [5], [6], and ECN has been proven to be a better way to deliver congestion signals and to exhibit better performance [4], [5], [7].

In most ECN implementations, when congestion happens, the congested router marks the incoming packet that just entered the queue. When the buffer is full or when a packet needs to be dropped, as in random early detection (RED), some implementations, such as the ns simulator [8], have the “drop-from-front” option, as suggested by Yin [9] and Lakshman [10]. A brief discussion of drop-from-front in RED can be found in [11]. However, for packet marking, these implementations still pick the incoming packet and not the front packet. We call this policy “mark-tail”.

In this paper, we propose a simple marking mechanism – the “mark-front” strategy. This strategy marks a packet when the packet is going to leave the queue and the queue length is greater than a predetermined threshold. The mark-front strategy is different from the current mark-tail policy in two ways. First, since the router marks the packet at the time when it is sent, and not at the time when the packet is received, a more up-to-date congestion signal is carried by the marked packet. Second, since the router marks the packet in the front of the queue and not the incoming packet, congestion signals do not undergo queueing delay as do the data packets. In this way, a faster congestion feedback is delivered to the source.

The implementation of this strategy is extremely simple. One only needs to move the marking action from the enqueue procedure to the dequeue procedure and choose the packet leaving the queue instead of the packet entering the queue.

We justify the mark-front strategy by studying its benefits. We find that, by providing faster congestion signals, the mark-front strategy reduces the buffer size requirement at the routers; it avoids packet losses and thus improves the link efficiency when the buffer size in routers is limited. Our simulations also show that mark-front strategy improves the fairness among old and new users and alleviates TCP's discrimination against connections with large round-trip time (RTT).

The mark-front strategy differs from the drop-from-front option in that, when packets are dropped, only implicit congestion feedback can be inferred from timeout or duplicate ACKs; when packets are marked, explicit and faster congestion feedback is delivered to the source.

Gibbons and Kelly [6] suggested a number of mechanisms for packet marking, such as “marking all the packets in the queue at the time of packet loss”, “marking every packet leaving the queue from the time of packet loss until the queue becomes empty”, and “marking packets randomly as they leave the queue with a probability so that later packets will not be lost”. Our mark-front strategy differs from these marking mechanisms in that it is a simple marking rule that faithfully reflects the up-to-date congestion status, while the mechanisms suggested by Gibbons and Kelly either do not reflect the correct congestion status or need sophisticated probability calculations for which no sound algorithm is known.

It is worth mentioning that the mark-front strategy is as effective in high speed networks as in low speed networks. Lakshman and Madhow [12] showed that the amount of drop-tail switches should be at least two to three times the bandwidth-delay product of the network in order for TCP to achieve decent performance and to avoid losses in the slow start phase. Our analysis in Section 4.3 reveals that, in the steady-state congestion avoidance phase, the queue size fluctuates from empty to one bandwidth-delay product. So the queueing delay experienced by packets when congestion happens is comparable to the fixed round-trip time.¹ Therefore, the mark-front strategy can save as much as a fixed round-trip time in congestion signal delay, independent of link speed.

This paper is organized as follows. In Section 2, we describe the assumptions for our analysis. The dynamics of queue growth with TCP window control is studied in Section 3. In Section 4, we compare the buffer size requirements of mark-front and mark-tail strategies. In Section 5, we explain why mark-front is fairer than mark-tail. The simulation results that verify our conclusions are presented in Section 6. In Section 7, we remove the assumptions made to facilitate the analysis and apply the mark-front strategy to the RED algorithm. Simulation results show that mark-front has advantages over mark-tail, as revealed by the analysis.