ALSketch: An adaptive learning-based sketch for accurate network measurement under dynamic traffic distribution

Abstract

Learning-based sketch improves accuracy and reduces the space overhead of traditional sketches by using machine learning to recognize input data patterns. However, most learning-based sketches adopts offline learning to train a learning model, which shows poor accuracy in measuring dynamic network traffic. To accommodate unpredictable changes in network traffic, it is urgent to study a learning-based sketch that not only supports timely updating recognition rules but also keeps accurate and efficient network traffic measurement. We propose a novel adaptive learning-based sketch (ALSketch) to accurately and efficiently measure dynamic network traffic with online machine learning. ALSketch incrementally constructs a learner to identify hot and cold flows and store them separately. Such a strategy reduces hash conflicts and improves estimated accuracy under limited memory usage. We perform extensive comparative experiments to verify the performance of ALSketch. Experimental results reveal that the average relative error of ALSketch is approximately 1.86∼5.25 times lower than that of traditional sketches under the same memory conditions.