OFBench: OpenFlow交換器效能測試方法

Abstract

OpenFlow 交換器已在現今網路佔有一定地位，隨之帶來的效能議題 也倍受矚目。為此，我們提出五項測試方法供驗證各家OpenFlow 交換器 效能評估，當中的測試指標包含action time、pipeline time、packet-in rate、 packet-out rate、pipeline gain 及timeout accuracy；此外，為了簡化測試流程， 我們設計一自動化測試系統OFBench，該系統為controller-agent 架構，並提 供測試項目的開發功能。實驗顯示，硬體交換器在timeout 的實作上較為不 足，其實際誤差約為20%，但在Packet-in 測試項目中，硬體交換器相對 於軟體交換器約有20 倍的效能差異，另外，在table pipeline 測試項目中， 硬體交換器在1Gbps 的流量下擁有約40 􀀀 60% 的pipeline gain 且其實際可 使用的table 數量僅有60 多張。而在實驗過程中，我們額外發現交換器在部 分功能實作上有3 項非預期的行為，其一，交換器在Apply-Action 功能並沒 有實作完全；其二，交換器在處理大量Packet-in 流量時，會導致斷線需重 新進行連線；最後，交換器在處理idle-timeout 時，會有未正確將該值重置， 導致flow entry 提早到期。

Currently, there are many OpenFlow switch products available in the market. The performance issues of OpenFlow protocol are drawing a lot of attentions. In this work, we propose five test cases to evaluate six performance metrics: action time, pipeline time, packet-in rate, packet-out rate, pipeline gain, and timeout accuracy. The switch can be evaluated based on these metrics. And we also propose the automatic test framework: OFBench, which is a controller-agent architecture allowing the development of test case based on high-level scripte language. The evaluation results reveal a 20% skewness in idle-timeout accuracy in the hardware switches. For the Packet-in rate, the hardware switches is capable of generating Packet-in frames in the rate ranging from 3000 to 7000 Packet-in per second. The rate is 20 times faster than that generated by the software switches. As for the test of the gain of table pipeline, the hardware switches reach 40 􀀀 60% pipeline gain under the 1Gbps traffic loading. And the switches only have around 60 usable tables. Furthermore, we observed three issues for switch during the testing. Firstly, the switches may not be well implemented on the design of Apply-Action instructions. Secondly, the switches suffer from random crashes with the high volume of bursty Packet-in traffic. Lastly, the timer of idle-timeout is not reset properly when the flow entry matched.