Online Elephant Flow Prediction for Load Balancing in Programmable Switch-Based DCN

In the data center network, traffic has a distinct heavy-tailed distribution characteristic, with the minority of throughput-sensitive elephant flows occupy most of the bandwidth and the majority of latency-sensitive mice flows require low latency. Therefore, it is very important to predict the network flow size and make a reasonable balanced scheduling. Currently, the traditional elephant flow detection schemes based on thresholds have poor accuracy and low granularity, while the intelligent detection schemes based on SDN has a certain flow scheduling response delay. For this reason, a two-stage online elephant flow prediction method for load balancing (OPLB) is proposed. Based on the programmable data plane, OPLB first pre-identifies the elephant flow by extracting the stateless features of the first packet of the network flow arriving at the switch. Secondly, the size of the elephant flow is predicted by extracting the features of the first ${n}$ packets of the flow. Finally, the detected elephant and mice flows are balanced to high throughput and low latency paths. Combined with the computing and storage capabilities of the programmable switch, the models and parameters in OPLB can be updated online by mapping the trained classification and prediction decision tree models to the matching-action pipelines of the programmable switch, thus achieving dynamic load balancing. We prototype OPLB in P4 software simulation environment and evaluate it with packet traces from the university data centers (UNI). The experiment shows that the accuracy of classification and prediction reached 89.3% when the proportion of elephant flow was 20%. At the same time, compared to the scheme that only uses single stage elephant flow prediction, OPLB reduces the amount of information collected by the switch by about 40% when the elephant flow proportion is 20%.