Self-adaptive Probabilistic Sampling for Elephant Flows Detection

Sampling traffic traces to collect data from Internet nodes offers a new approach to reduce the traffic amount of sampling and the memory depletion. However, the loss of accuracy in traffic detection may be large due to the impertinent sampling probability. Many elephant flows detection methods suffer the high time consumption and the poor detection results from the empirical sampling. In this paper, the self-adaptive sampling method for elephant flow detection is investigated based on the heavy-tailed distribution of Internet. Specifically, the SaPS (Self-adaptive Probabilistic Sampling) algorithm is proposed to capture the characteristics of heavy-tailed flows through periodically calculating the kurtosis of tailedness for the flows. We employ this algorithm to provide simplified but representative samples to four well-known detection algorithms. The results of extensive simulations show that our sampling algorithm achieves the high performance in terms of time and memory consumption while maintaining a high accuracy through dynamically adjusting sampling probability for elephant flows detection.