Noise can help: accurate and efficient per-flow latency measurement without packet probing and time stamping

Abstract

With the growth in number and significance of the emerging applications that require extremely low latencies, network operators are facing increasing need to perform latency measurement on per-flow basis for network monitoring and troubleshooting. In this paper, we propose COLATE, the first per-flow latency measurement scheme that requires no probe packets and time stamping. Given a set of observation points, COLATE records packet timing information at each point so that later for any two points, it can accurately estimate the average and standard deviation of the latencies experienced by the packets of any flow in passing the two points. The key idea is that when recording packet timing information, COLATE purposely allows noise to be introduced for minimizing storage space, and when querying the latency of a target flow, COLATE uses statistical techniques to denoise and obtain an accurate latency estimate. COLATE is designed to be efficiently implementable on network middleboxes. In terms of processing overhead, COLATE performs only one hash and one memory update per packet. In terms of storage space, COLATE uses less than 0.1 bit per packet, which means that, on a backbone link with about half a million packets per second, using a 256GB drive, COLATE can accumulate time stamps of packets traversing the link for over 1.5 years. We evaluated COLATE using three real traffic traces that include a backbone traffic trace, an enterprise network traffic trace, and a data center traffic trace. Results show that COLATE always achieves the required reliability for any given confidence interval.