Analysis and evaluation of the TCP/IP protocol stack of LINUX

With the rapid progress of high-speed network technology, the communication speed between hosts will reach Gbit/s. So it is important to investigate the traditional TCP/IP protocol stack to find out whether it is suitable for high-speed communication and can provide QoS or not. Based on this consideration, the LINUX TCP/IP stack's structure, the buffer management, the procedures for sending and receiving data, and the interaction with the OS are analyzed in this paper. A probing node method is proposed to investigate the internal structure of the stack. With this method, the throughput of the stack, the time consuming of each layer and the queuing of the IP layer at high speed are tested. Our experiments show that the throughput of the TCP/IP stack of LINUX can reach 360 Mbit/s at the sending side and 150 Mbit/s when the local host device is used for bi-directional operation, the protocol stack can he improved to support higher speed; the percentages of the various costs of the TCP/IP protocol suite have changed greatly, zero-copy and zero-scan techniques cannot improve the performance dramatically; there is no queuing at the receiving side of the IP layer, so FCFS (first come first served scheduler) is enough, but a more sophisticated packet scheduler such as WFQ (weighted fair queuing) is needed at the sending layer of the IP layer to provide QoS in the future.