PI controller is based on TCP/AQM cybernetics model proposed by Mirsa et al.
It uses control theory to study active queue management algorithm. In PI controller, the parameters are fixed and the performance is poor in high-speed network, which can not meet the design objectives of AQM. Therefore, an enhanced PI control algorithm, EPI, is proposed in this paper. The average queue length is used to adjust the packet loss probability online, so that PI can meet the performance requirements of high-speed networks. The simulation results show that the comprehensive performance of EPI is better than that of PI in high-speed network. Research shows that the current AQM algorithm is far from meeting the needs of high-speed networks, and an effective AQM algorithm must be designed to solve this increasingly serious problem. Although RED algorithm is recommended as the only candidate for active queue management by RFC2309, there are still many imperfections in RED algorithm itself, mainly in stability and fairness.
In order to improve and perfect the deficiencies of RED algorithm, there are many variants of RED algorithm and new AQM algorithm.
Hollott et al. proposed PI controller based on the linearization of the established cybernetics model.
PI controller has smaller queue jitter than RED algorithm. But in PI controller, the parameters are fixed, which results in poor performance of PI controller in high-speed network.
Reference [2] regards PI controller as an ADALINE neural network with two input variables. LMS algorithm is used to adjust the proportion factor and integral factor online. Reference [3] dynamically adjusts the parameters of PI algorithm on the basis of PI controller. In reference [4], a distributed dynamic bandwidth allocation algorithm is proposed. These algorithms have some adaptability to dynamic environment, thermostatic element but the calculation is complex, which increases the burden of routers.
Therefore, this paper proposes a strengthened PI controller, which uses average queue length to adjust the packet loss probability online, so that PI can meet the performance requirements of high-speed network. In order to investigate the performance of EPI controller in high-speed network, the author uses NS simulator to simulate.
The simulation environment is shown in Figure 2. The parameters of EPI are as follows: delta = 0.25, alpha = 0.025, and beta = 0.005.
These parameters are empirical values and should be set according to different environments. Figure 3 shows the bottleneck link bandwidth utilization in N1 nodes when EPI and PI are used separately. It can be seen from the figure that the bandwidth utilization of EPI is higher and more stable than that of PI.
The average bandwidth utilization of EPI is 88.3%, which is 12.84 percentage points higher than that of 75.46% of PI. Figure 4 shows the packet loss rate of N1 when EPI and PI are used in N1 node respectively. From the figure, it can be seen that the packet loss rate of PI has been around 0.0002, although sometimes unnecessary.
Combined with Figure 3 and Figure 4, it can be seen that PI controls the queues in N1 node too radically, resulting in a reduction in bandwidth utilization. Although the packet loss rate of EPI is slightly higher, it guarantees a higher bandwidth utilization. Fig. 5 shows the queue length of N1 node when EPI and PI are used respectively.
It can be seen from the graph that PI can control the queue very small, but at the expense of bandwidth utilization, but the queue length of EPI is very stable and can basically control the queue length within the target value of 1000 packets. Aiming at the disadvantage of PI controller with fixed parameters and poor performance in high-speed network, this paper proposes an EPI controller which uses average queue length to adjust packet loss probability online, which enhances the adaptability of PI controller to high-speed network environment. The simulation results show that the comprehensive performance of EPI in high-speed network is better than that of PI.