Machine Learning based Flow Classification in DCNs using P4 Switches

This paper deals with classifying flows in data center networks, primarily based on the flows' volume of traffic and duration. Flows are typically classified as long-lived flow or short-lived flow. Long-lived flows throttle the short-lived flows and should be classified at the earliest to select a different path in the network for them. The objectives of the proposed classification scheme are: (i) to support more than two flow classes (three in this paper), (ii) to achieve early classification by observing the first few packets in the flow, (iii) to achieve classification using ML techniques implemented in a programmable data plane switch using the Programming Protocol-independent Packet Processors (P4) language. Our contribution includes an improved hash-and-store algorithm for flow classification. The ML technique considered is Decision Tree, since it can be efficiently implemented in a P4 environment. The techniques have been evaluated using simulation-generated data implemented in a mininet emulator environment and classification accuracy results obtained. Two existing schemes, HashPipe and IdeaFix have also been implemented for comparison. The results show that the proposed scheme can classify a flow within 3 MB of the flow size when we consider more than one feature to classify the flows. This outperforms the existing threshold-based schemes by classifying flows, 3 times faster.

Automated summary

This paper presents a scheme for classifying flows in data center networks based on the volume and duration of traffic, using ML techniques implemented in programmable data plane switches. The proposed scheme outperforms existing threshold-based schemes in terms of flow classification accuracy and speed.