優化裸機商業交換器處理 OpenFlow 訊息的效能

Abstract

當一台OpenFlow交換機的晶片收到一個沒有在flow table中匹配任何一條entry的封包，這個封包會從交換機的晶片被複製到交換機的中央處理器去做處理。中央處理器會把這個封包以Packet-In訊息的方式送到SDN控制器，然後SDN控制器又會將這個Packet-In訊息轉成Packet-Out訊息再送回交換機。中央處理器收到後，解析Packet-Out header、取得output port資訊，並將此封包從output port送出去。這一連串的4個操作會頻繁地發生，而且這些操作的執行效能會大幅影響到整個SDN網路的運作。在這篇碩士論文中，我們會以這4個操作為一組做最佳化，以提升throughput和降低latency為目標做效能改進。實驗平台是一台支援OF-DPA的48 port x 10 Gbps (port 頻寬)的SDN裸機商業交換機。實驗結果顯示我們的方法可以提升throughput 2.5到5700倍，並降低latency 2到167倍。

When the switching ASIC of an OpenFlow switch receives a packet that cannot match any entry in its flow table, the packet needs to be copied to the switch CPU, then the switch CPU transfers it to the SDN controller via a Packet-In message, then the SDN controller converts the Packet-In message to Packet-Out message and transfers back to the switch CPU, then the switch CPU forwards the packet out of an output port specified in the Packet-Out message. The group of the four consecutive operations occurs very frequently and their performance greatly affect the performance of the overall SDN network. In this paper, we optimized the throughput and latency performance of such group operations in a 48 port x 10 Gbps (port bandwidth) SDN bare metal commodity switch that supports the OF-DPA mechanism. Our experimental results show that our design and implementation can speed up the throughput by a factor of between 2.5 and 5,700 and reduce the latency by a factor of between 2 and 167.