分波多工光分封交換機在自我相似訊務下之效能評估

Abstract

隨著日益增加的網際網路訊務流量造成的頻寬需求，分波多工技術結合全光分封交換機可提供大量頻寬及高速交換能力，因此，成為了建構下一世代光纖網路最有潛力的解決方案。在本論文中，我們探討了分波多工光分封交換機的排隊效能及不同的緩衝器排程管理策略，並且藉由馬可夫近似模型提出分析分波多工光分封交換機在自我相似性訊務下之可行方法。首先，在第二章中，我們探討了前授式及回授式分波多工光分封交換機在時槽式同步及可變封包長度之自我相似訊務下的排隊效能分析，並說明回授式光交換機的優點及各種系統參數及訊務特性參數對於排隊效能之影響。在第三章中，我們首先說明在可變長度封包訊務下，在緩衝器中會造成空隙，此空隙問題會造成交換效能的下降。傳統的空隙填塞排程演算法雖然可以改善此問題，但是其運算複雜度太高以致於無法實現在高速分波多工光交換機上。為此，我們提出了一個新且有效的空隙填塞排程演算法應用於分波多工光分封交換機上，藉由數學分析以及模擬結果得知，此演算法可大幅降低傳統演算法之運算複雜度並且仍然維持良好的交換效能。在第四章中，我們首先分析各種封包排程策略之優劣，並且探討其在出口端緩衝器上以及在回授式共用緩衝器上的影響，最後提出能達到最低封包漏失率的最佳排程策略。藉由運用此簡單的排程策略，可以在不增加多少運算複雜度的情況下，大幅改善回授式分波多工光分封交換機的效能。在第五章中，我們延伸之前的單一伺服器馬可夫訊務模型研究到多重伺服器佇列及單一伺服器優先權管理佇列之排隊效能分析。我們提出了馬可夫訊務模型近似自我相似訊務在兩種佇列上的數學解，並且用模擬證明其準確性。藉由與具有二階自我相似特性的分數式布朗訊務之模擬結果比較，可證明此研究結果可以應用於分析前授式分波多工光交換機在自我訊務下之交換成效及具有優先權管理之光交換機排隊效能。最後，我們在第六章中總結本篇論文，並且說明未來可延伸的研究工作及方向。

As the current network evolves, the traffic increases so rapidly that high speed optical network becomes the most imperative issue to provide the next-generation Internet service. The all-optical switches are especially the most pending problem to be solved. In this thesis, we first compare the switching performance between two kinds of optical packet switches handling variable length packets and discuss their pros and cons. We then propose a new void filling algorithms that could greatly reduce the computation complexity and still maintain satisfactory performance. Next, we analyze the effects of various packet scheduling algorithms on the optical packet switches and recommend the best one to be used. At last, we propose an efficient approximation for analyzing the switching performance of wavelength division multiplexing (WDM) optical packet switches handling self-similar traffic. According to the same philosophy, we further analytically derive the approximate analysis of optical packet switches employing space priority mechanism (specifically, partial buffer sharing mechanism) under self-similar traffic input. The essential abstract of this thesis is depicted as follows. In Chapter 1, we first introduce the motivation and the background of the re-searches in this thesis. We believe that all-optical network is the ultimate and the best solution in future to handle the ever-increasing bandwidth demands of the next generation Internet. The switching and routing capability of the network nodes are especially crucial issues to be solved. However, the conventional electronic switches could not provide such high-speed switching capability. The transmission rate of optical fiber utilizing WDM technology has reached THz in commercials, and the all-optical switches could further exploit the benefit of optical fiber communication. Therefore, the switching performance of optical packet switches combining with the dense wavelength division multiplexing (DWDM) technology that can multiply the bandwidth of a single fiber by exploit the wavelength domains is worth to be inves-tigated. On the other hand, the WAN and LAN traffic both show the self-similarity, which degrades the switching performance severely. The traditional queueing model could not analyze the switching performance under self-similar traffic very well. Thus, we also employ the generalized variance-based Markovian fitting model for self-similar traffic in order to provide a feasible analytical method for evaluating the switching performance under self-similar traffic. In Chapter 2, we want to have a thorough comparison of the feed-forward (FF) and feedback (FB) type WDM optical packet switches under variable-packet-length self-similar traffic. We first introduce the operation and the architectures of the FF and FB type WDM optical packet switches. Next, we program the discrete event time simulator in C++ language in order to analyze the switching performance of the FF and FB type WDM optical packet switches. The simulation is under various system dimensions and traffic parameters. Based on the simulation results, it is observed that the FF type could not handle the self-similar traffic very well even under less bursty traffic. The FB type optical packet switches with commonly shared re-circulated buffers could provide extra buffer space for collided packets and smooth the bursty traffic. Thus, the FB type WDM optical packet switches is better than FF type WDM optical packet switches, and we will adopt the FB type switches in the following analysis in this thesis. On the other hand, it is also observed that the packet loss performance of FB type WDM optical packet switches depends on the parameters, and we need to carefully choose the correct dimension in order to obtain the optimal switching performance. In Chapter 3, we propose a newly efficient void filling algorithm for FB type WDM optical packet switches. We first explain the formation of the voids in WDM optical packet switches that degrades the packet loss performance severely. The traditional void filling algorithm is shown to be effective to improve the switching performance, however, it takes too much computation complexity to be implemented on the high-speed WDM optical packet switches. Hence, we propose the new void filling algorithm which reduces the number of voids to be stored and processed. We analyze the computation complexity and computation time and discover that the newly proposed void filling algorithm could hugely decrease the computation complexity. Moreover, we also perform simulation of buffer utilization and switching performance of these two void filling algorithms on FF and FB type WDM optical packet switches. From the simulation results, these two void filling algorithms show approximate results. Therefore, we conclude that the new void filling algorithm could indeed reduce the computation complexity while producing satisfying switching performance, and it is more suitable to be implemented on FB type WDM optical packet switches. In Chapter 4, we discuss the packet scheduling algorithms applied to the FB type WDM optical packet switches in order to further improve the switching performance. The packet scheduling algorithm determines which packet has prior order to enter the virtual wavelength queues. For the analysis in this chapter, we also adopt the better wavelength allocation strategy, joining the virtual wavelength queue with the minimum queue length strategy, to obtain the optimal packet loss performance. For FB type WDM optical packet switches, five different packet scheduling algorithms can be utilized. In this chapter, we first examine that the advantage and disadvantage of every packet scheduling algorithm, and discuss the packet loss performance of these algorithms. Next, we discuss the effectiveness of these algorithms on primary output buffers and re-circulated buffers, and it is shown that the packet scheduling algorithm has stronger in°uence on packet loss performance when applied to primary output buffers. Finally, we try different combination of packet scheduling algorithms and find out the optimal strategy with minimum packet loss performance. Thus, we could use appropriate packet scheduling algorithms in order to totally exploit the benefits of FB type WDM optical packet switches. In Chapter 5, we extend the previous research of generally variance-based Markov-ian fitting for self-similar traffic to the multiple-server queues and space priority queues on a single-server queue. We first introduce the procedure of analysis of these two cases and prove that the numerical results is valid by MMPP simulations. Next, we demonstrate the MMPP numerical results could also fit the queueing performance of the queues under self-similar traffic input. Based on the MMPP numerical results, MMPP simulation results, and self-similar traffic results, we could confirm that this fitting method actually conforms to the self-similar traffic, and we provide a very useful method to handle the queueing performance under self-similar traffic. At last, we summarize the research work that we have completed and conclude this thesis in Chapter 6. However, there are still some important issues on WDM optical packet switches need to be investigated. These issues are presented in the future work section in Chapter 6.