非同步模式網路最佳化及模糊適應預估之流量控制

Abstract

非同步模式網路架構，在目前高速寬頻網路中是相當廣泛的網路傳送技術並扮演極重要的角色，極大部分即時及非即時交通流量，在有線及無線通訊網路中會匯集至非同步模式網路，本論文提供三種方法來整合非同步模式網路上各式各樣型式的交通流量，以避免擁塞，並達成高度頻寬使用率及保證服務品質。 於第二章中，首先提出一控制理論的演算法則，調整可用位元速率(ABR)的頻寬，以處理網路擁塞問題。本法則藉由史密斯預估器克服迴路傳輸延遲及確保回饋迴路穩定度，而可用位元速率使用者的頻寬，則基於最小細胞速率(MCR)要求，將網路頻寬扣除服務品質要求嚴格的固定位元速率(CBR)及變動位元速率 (VBR)頻寬總合再加以平均分攤，進而採用簡單及最佳化的 H∞ 流量控制器來降低佇列延遲(queuing delay)及其變動量(variation)，以避免傳送細胞遺失(cell loss)。由於最佳化的H∞ 流量控制器沒有解析解(analytic solution)，因此運用基因法則(genetic algorithm)來尋找最佳控制參數。 於第三章中，則運用流量非線性時變的特性並提出最佳化流量模糊估測與ABR流量適應控制，對於一個節點(node)之輸入流量(包含可控制的ABR流量與不可控制的VBR/CBR流量)以及佇列長度(queue length)，以模糊自動遞迴移動平均加上外加訊號(Fuzzy ARMAX)的模糊流量模式表示。為了克服擁塞控制中傳輸延遲所造成的困難，上述模糊流量模式將轉換成一個等效的預估式模糊流量模式，進而用適應控制方法，估測預估式模糊流量模式中的參數，以達到精確的網路流量估測，並對佇列作最小變動量(minimum variance)控制。 最後於第四章中提出對於時變、突暴型(burst)及長期相關的變動速率(VBR)服務中MPEG視訊流量作估測，我們將MPEG視訊流量中的長週期、短週期與線性趨勢(linear trend) 以狀態空間隨機動態模式 (state-space stochastic dynamic model)表示，同時MPEG視訊流量估測問題可轉為一個隨機狀態空間模式的狀態估測問題。由於此模式的雜訊統計特性為未知，我們採用H2/H∞ 濾波器法則做狀態估測，經由模擬結果及比較時變的H∞ 濾波方案及時間延遲神經網路(TDNN)方法，我們發現本文所提出的非時變H2/H∞ 濾波法是基於成效及計算成本考量下的不錯的折衷方案。

ATM network is the most popular transport technology and becomes more and more important on broadband network recently. Most real time or non real-time traffic of wired or wireless network will be integrated or internetworked to ATM networks. To integrate various types of traffic, avoid traffic congestion and ultimately gain the high bandwidth utilization on ATM networks, three issues are considered in this dissertation. First, a control-theoretic algorithm to deal with the congestion control problem is presented to adjust the ABR source rates. With the aid of Smith predictor, system can overcome the effect of multiple propagation delays and ensures the stability of the feedback loop. The available ABR bandwidth left by quality of service (QoS) constrained traffic (CBR+VBR) is fully and fairly utilized in the sense of MCR plus equal share. This fairness property holds even if there are uncertainties or measurement errors of the parameters. Further, a simple optimal H∞ controller is proposed to minimize the variance of the queue occupancy and avoid cell loss. Since no closed form solution can be found for the optimal H∞ traffic controller, an effective genetic algorithm to find the control parameters is also provided on the algorithm. Secondly, another scheme is proposed to exploit the nonlinear time-varying property of network traffic and develop an adaptive optimal flow control for adjusting the ABR source rates. The incoming traffic flow and the queuing dynamics are modeled by a fuzzy autoregressive moving-average model with auxiliary input (fuzzy ARMAX process), with the traffic flow from uncontrolled sources (i.e., cross traffic) being described as external disturbances. In order to overcome the difficulty of the transmission delay in the design of congestion control, the fuzzy traffic model is translated to an equivalent fuzzy predictive traffic model. A fuzzy adaptive flow control scheme is proposed to avoid congestion at high utilization while maintaining good quality of service. By use of fuzzy adaptive prediction technique, the most difficulties in congestion control design due to nonlinearity, time-varying characteristics, and large propagation delay can be overcome by the proposed adaptive traffic control method. Furthermore, an optimal adaptive fuzzy predictive control scheme can effectively process nonlinear traffic model and achieve the goals of maintaining QoS, high network utilization and minimum buffer variation. A comparative evaluation is also given to show the superiority of the proposed method. Finally, a prediction scheme is brought up to predict the burst and long-range dependent traffic for real-time VBR MPEG video. The trend and periodicity characteristics of MPEG video traffic are fully captured by a proposed state-space stochastic dynamic model, which includes traffic parameters in the state vector, to improve the accuracy of prediction. As the statistics of the underlying processes are either unavailable or uncertain in real-time applications, an H2/H∞ filtering algorithm is proposed to estimate traffic parameters for long-range prediction. Unlike previous prediction schemes, which predict I, P, B frames separately, the proposed scheme is simplified to predict the composite MPEG video traffic. Simulation results based on real MPEG traffic data show that the time-varying trend, the periodic components, and the long-range dependence property can be splendidly predicted and captured by the proposed method. Compared with the time-varying H∞ filter and the TDNN algorithm, the proposed time-invariant H2/H∞ filtering algorithm is a good compromise based on performance and computation cost.