流量矩陣量測及軟體定義網路多路徑遶送

Abstract

此論文提出了在軟體定義網路中實施多路徑遶送及流量矩陣(Traffic Matrix)量測的方法。我們利用OpenFlow中定義的Group Table支援多路徑遶送，其中使用的是select類型，選擇不同的通訊埠實施多路徑遶送。要支援多路徑遶送在Flow Table中只需對目的地IP地址，但為了更有效率量測網路流量矩陣，我們針對每一流量來源子網路及其目的子網路的組合，加入一比對來源IP地址及目的IP地址的flow entry；此一流量表內容比只比對目的地IP地址的內容有較高的優先。我們可以直接利用此類新增流量表內容的計數器計算出流量矩陣。 此外，在OpenFlow定義下select類型的權重(weight)參數並不明確，為了實作軟體定義網路下的多路徑遶送功能，我們重新定義權重參數的比例應等於各路徑遶送流量的比例。在OpenvSwitch下的權重參數定義並非如此，因此我們提出了三個要使權重參數與實際流量比例相等的議題，並提出方法去解決這些議題，重新修改了OpenvSwitch的功能，使我們可以準確的實作多路徑遶送功能。我們在修改後的OpenvSwitch可以支援多路徑遶送也可以達到流量權重與權重參數相同的目的。此外驗證流量矩陣量測，確實可量測得流量矩陣。 我們的實驗結果中也指出OpenvSwitch中Flow table的byte計數器每次更新為0.45sec，這代表著流量矩陣測量時可能會因為這間隔導致不正確。因此應用在網路工程技術上的測量及控制可能不能與我們預期的設計一樣迅速。此外在未來研究中需要平穩的流量比例來驗證遶送路徑更動做小化。

This thesis presents methods to implementing multi-path routing and measuring network traffic matrix. We used the functions of Group Table defined in OpenFlow specifications; Group tables of type select were used to route traffic with the same destination IP subnet to different ports of a OpenFlow switch. To support multi-path routing, we only need to compare the destination IP addresses. However, to measure the traffic matrix more efficiently, we added an additional flow entry for each pair of a source IP subnet and a destination IP subnet. This flow entry has a higher priority than the flow entry that compares only the destination IP address. The byte counters of the added flow entries can be used to measure the network flow matrix. In addition, OpenFlow does not specifically define how the weight parametes used in Group table of type select should be implemented. We believe the weight ratios should be equal to the traffic ratio. However, the current implementation of OpenvSwitch the two ratio are not equal, and no efforts were spent to ensure that the traffic ratio to be stable for the same weight ratio. We investigate three design issues in improving the implementation of Group Table of type select, and implemented our solutions for OpenvSwitch. Our experiment results indicate that using our modified OpenvSwitch, we can support multi-path routing and maintain the flow ratio equal to the weight ratio. In addition, the traffic matrix can be measured. Our experiment results also indicated that the byte counters of the flow tables on OpenVswitch only be updated about every 4.5 secs. This means that taffic matrix measured for a short interval may not be accurate. As a result, the measurements and control cycle used for network traffic engineering cannot be as short as one may desire. In addition, further investigations are needed to maintain a stable traffic ratio for multi-path routing while minimizing the re-routing traffic.