输入排队的交换结构中的TCP拥塞控制

TCP Congestion Control in Input-queued Crossbar Switch

交换结构是高速路由器不可缺少的重要组件之一,随着高速网络传输和操作的媒质从电域变化到光域,光缓存和光控制技术的局限性使得交换结构极有可能采用简单输入排队的工作模式。在简单输入排队的交换结构中,TCP的拥塞控制存在着不同寻常的新问题,本文首先指出在交换结构环境下,单靠传统的队列管理机制无法兼顾交换结构的吞吐量和TCP公平性。分析了队列管理机制之所以失效的原因,在于它无法同时克服队头堵塞和TCP流对交换结构端口的"不公平争抢"问题。接着提出一种把公平的启发式仲裁和随机提前丢头队列管理机制(hFs&rEDF)结合起来的算法,能够兼顾交换结构吞吐量和TCP公平性。仿真实验表明hFs&rEDF算法可以获得75%以上的吞吐量和接近于1的公平性。

Switches are necessary and important components of high-speed routers. The transportation and operation media in high-speed networks being transfered from the electronic to optical domain, it is possible that input-queued switches, where only one queue is at each input port with finite buffer space, are used due to limitation of the optical buffer and control technology. The paper presents that in such an input-queued switch while TCP protocol is employed for end-to-end congestion control, the congestion control problem is interesting. Firstly we found that it was impossible to achieve high throughput and fairness among TCP flows at the same time only by queue management. We also analyzed why queue management mechanisms did not work, the reason was that it couldn＇t resolve the HOL problem and unfair contention among TCP flows for output ports of switches at the same time. Then we proposed a scheme of hFS＆rEDF, which provided high throughput and fairness by combining heuristic fair arbitrating （hFS） with queue management mechanism of early dropping front randomly （rEDF）. Some simulation results proved that this scheme could achieve above 75% throughput and fairness near 1.