From: Michelle Liu (liujing@u.washington.edu)
Date: Sun Oct 17 2004 - 22:14:04 PDT
Review of "Congestion Avoidance and Control"
Jing Liu
This paper talks about mechanisms of TCP congestion control and avoidance in networks. Since there is no status information provided by the networks to the hosts, thus the hosts can not have an overall picture of network traffics and also some hosts could be mis-behaving, congestion could happen in networks.
The author adds several new algorithms into TCP. In this paper, the author focuses on explaining the mechanisms of slow-start, round-trip-time (RTT) variance estimation and dynamically adjusting the congestion window. The rationale behind them is also stated clearly. For instance, slow-start is to avoid overshooting traffic of hosts in the beginning and to avoid lengthy initial delay as well. Dynamically varying the window size is necessary to avoid either unnecessary packet loss or underutilize the bandwidth of the network. Addictive increase and multiplicative decrease of the window size is a quite good mechanism because of the fact that it is very easy to drive the net into saturation but hard for the net to recover.
From the performance charts, we can see that those mechanisms work quite well. The continuous retransmissions are much less than before.
There are some weaknesses of the proposed algorithms. First, it is not perfect for the mechanism of addictive increase and multiplicative decrease of the window size. When congestion happens, numerous hosts will lose packets and all of them will drop the window size dramatically at the same time. The utilization of network bandwidth at that time could be very low. Second, it takes some time to increase the utilization of bandwidth, since the window size is increased addictively. Third, there must be a period of oscillation of TCP traffic in the networks before each TCP finds the correct window size. Fourth, the slow-start starts from one packet to increase the bandwidth utilization. If the link bandwidth is very high, it takes numerous RTTs to reach the full utilization of the bandwidth following a burst of congestion. In addition, since TCP uses dropping packets to indicate congestion, misbehaved sources can cause congestions all the time and affect the whole network performance.
As for the future work, one thing could be done is to improve the mechanisms with regard to the above problems. The other thing is that since the TCP notices congestion happening by receiving the signals of packet loss, the gateways must provide 'congestion detection' algorithms or making decision on which packets should be drop in the queues. The decision influences fair allocation of bandwidth. Somehow, some QoS could be provided by a good 'congestion detection' algorithm.
This paper is very useful today. We use TCP congestion controls in today's Internet and the network performance has been improved a lot before we didn't use it.
This archive was generated by hypermail 2.1.6 : Sun Oct 17 2004 - 22:14:05 PDT