支援加權公平佇列多頻道並預複製遷移之都會型網路研究

在都會型網路Metropolitan Area Network (MAN)中，網路範圍定義是介於廣域網路與區域網路之間。不一樣的應用在都會型網路裡，其中一項眾所皆知的即是有線電視(Cable Television)。它遵從著有線電纜資料服務介面規範Data Over Cable Service Interface Specification (DOCSIS)。此規範是由美國有線電視實驗室(Cable Television Laboratories, CableLabs)開發完成，且它支援雙向同時傳輸並應用網際網路協定(Internet Protocol, IP)。 “加權公平排程機制於有線電纜資料服務介面多頻道下行之研究”的主要目的是在有線電纜資料服務介面規範中除了支援服務品質保證(Quality of Service, QoS)技術與頻道捆綁(Channel Bonding)技術的同時, 再加入加權公平排程機制(Weighted Fair Queueing, WFQ), 使得整體服務兼顧品質保證與公平性，換句話說，如何兼顧這兩種特性, 是我們研究的課題。規範提供了五種服務品質保證與頻道捆綁技術機制, 透過纜線數據機(Cable Modem, CM)經由上行(Upstream)提出服務品質保證需求給纜線數據機終端系統(Cable Modem Termination System, CMTS), 纜線數據機終端系統來做相對應的品質服務排程, 在應用中, 最常見運用的排程機制為先來先處理排程(First-Come First-Served, FCFS Queue)與優先等級排程(Priority Queue), 在這些排程機制下, 先來先處理排程使得網路延遲時間整體加長, 優先等級排程也使得網路可能因為優先權過低的封包, 在佇列中一直無法被系統處理到, 因此, 我們於研究中, 加入其加權公平排程機制之方法, 使得封包在分配排程上獲得較公平, 且不會因為優先權過低而發生無限期的阻塞(Indefinite Blocking)或飢餓(Starvation)的問題。在都會型網路, 還有另一個課題值得研究, 也就是雲端運算。雲端運算是最近很普遍受歡迎的應用，最近，雲端進而也延伸了另一種分散式運算架構的邊緣運算。DOCSIS也不例外，越來越多的網路服務業者提供如此便利的服務。一個原因為虛擬化技術的應用。它可以提供使用者於一台伺服器內架設很多虛擬機，便於各種資料的動態分配。並且，當任何一台伺服器可能因為節能省電需要進入睡眠模式，這時虛擬機將遷移至另一台伺服器繼續服務，使用者是不會有感覺中斷。這是仰賴於即時遷移(Live Migration)技術。 “運用賽局理論期望值對預複製預測之研究”會介紹到透過記憶體之修改機率預測，來決定是否記憶體在下一次修改時，即時遷移已不具效益，決定是否應該直接停止即時遷移，進入停機複製模式。在方法中我們將避免不必要的轉移減少整體遷移需要時間。之前的研究採取別種方法來預測記憶體變更的方法，但是一定要統計幾次疊代的歷史變化，才能準確預測記憶體修改的機率。在本研究中，我們將採用賽局理論模型，有效地減少預測髒頁(Dirty Page)數量來確定是否進入停止和直接複製階段，從而節省遷移所需的時間。

關鍵字

加權公平排程；多頻道下行；賽局理論；預複製

並列摘要

In Metropolitan Area Network (MAN), the network area size is defined between Wide Area Network (WAN) and Local Area Network (LAN). Different applications exist in MAN, one well-known application of them is Cable Television. The application complies with Data Over Cable Service Interface Specification (DOCSIS). DOCSIS was developed by Cable Television Laboratories (CableLabs). In DOCSIS network, it allows transparent Internet Protocol (IP) traffic in the system communication as well as a useful property of bidirectional transmission simultaneously. The aim of "A Weighted Fair Queueing (WFQ) Management on DOCSIS Multi-downstream Channels" is to develop a WFQ mechanism into DOCSIS network to effectively supply the downstream services. The work is based on Quality of Service (QoS) and Channel Bonding techniques, then implementing the mechanism of WFQ, overall service quality and fair distribution are considered. Totally five QoS models have been supported in DOCSIS. When one Cable Modem (CM) requests that support of QoS. At the end, CMTS will depend on the QoS request from CM to provide the relative service. With the service queues in pipes, general queues are First-Come First-Served (FCFS) Queue and Priority Queue, However, two defects exist. First, the total large latency will be obtained under FCFS Queue, that is because every packet is handled in queue one-by-one. Priority property is not supported in FCFS queue. Second, each packet in Priority Queue will be set with the order in priority, the packet with lower priority will be postponed when a packet with higher priority is sent at the same time. Even though an application of the priority queue can decrease the total latency compared to FCFS model, that also bring other issues, such as Indefinite Blocking or Starvation. According to above, we propose an approach of WFQ implementation in CMTS for solving the defects of FCFS and Priority Queue. The approach is not only to solve Indefinite Blocking or Starvation issue but also to make sure the lower latency is obtained based on the WFQ implementation. In MAN, the application of Cloud Computing is another popular technology in recent years. Recently, it became the approach further as to be distributed on Edge Computing. That idea did not exclude DOCSIS. More and more network applications provided clients more convenient experiences with the services. One of reasons is virtualization technology in Cloud Computing is applied, it improves the usage on the server and includes a characteristic dynamic data assignment. In Addition, any of servers for power-saving that entering Sleep Mode, which data needs to be migrated to another server for continuously operating, the user is not aware that the service has interrupted. That is because a technology of Live Migration will quickly backup the remaining data from the original server to another server. “An Expected Value Prediction of Game Theory for Pre-copy” introduces that is predicting the probability of memory modification to determine the dirty page whether to go the Stop-and-copy phase in the technology of live migration. This method can avoid data being unnecessary transferred over the network and reduce total time during the migration. The previous studies adopted other methods to predict the probability of memory modification. But it must be statistics with the historical changes at several iterations, to be able to accurately predict the probability of memory modification. In this study, we are going to adopt Game Theory model that effectively reduces the predicted numbers to determine the dirty page whether to go Stop-and-copy phase, thus saving time required for live migration.