Scalable and Efficient Diagnosis for 5G Data Center Network Traffic

Abstract

Data center networks (DCNs) for 5G are expected to support a large number of different bandwidth-hungry applications with exploding data, such as real-time search and data analysis. As a result, significant challenges are imposed to identify the cause of link congestion between any pair of switch ports that may severely damage the overall network performance. Generally, it is expected that the granularity of the flow monitoring to diagnose network congestion in 5G DCNs needs to be down to the flow level on a physical port of a switch in real time with high-estimation accuracy, low-computational complexity, and good scalability. In this paper, motivated by a comprehensive study of a real DCN trace, we propose two sketch-based algorithms, called α-conservative update (CU) and P(d)-CU, based on the existing CU approach. α-CU adds no extra implementation cost to the traditional CU, but successfully trades off the achieved error with time complexity. P(d)-CU fully considers the amount of skew for different types of network services to aggregate traffic statistics of each type of network traffic at an individual, horizontally partitioned sketch. We also introduce a way to produce the real-time moving average of the reported results. By theoretical analysis and sufficient experimental results on a real DCN trace, we extensively evaluate the proposed and existing algorithms on their error performance, recall, space cost, and time complexity.