Estimating TCP Latency Approximately with Passive Measurements
- Cite this paper as:
- Gangam S., Chandrashekar J., Cunha Í., Kurose J. (2013) Estimating TCP Latency Approximately with Passive Measurements. In: Roughan M., Chang R. (eds) Passive and Active Measurement. PAM 2013. Lecture Notes in Computer Science, vol 7799. Springer, Berlin, Heidelberg
Estimating per-flow performance characteristics such as latency, loss, and jitter from a location other than the connection end-points can help locate performance problems affecting end-to-end flows. However, doing this accurately in real-time is challenging and requires tracking extensive amounts of TCP state and is thus infeasible on nodes that process large volumes of traffic. In this paper, we propose an approximate and scalable method to estimate TCP flow latency in the network. Our method scales with the number of flows by keeping approximate TCP state in a compressed, probabilistic data structure that requires less memory and compute, but sacrifices a small amount of accuracy. We validate our method using backbone link traces and compare it against an exact, baseline approach. In our approximate method, 99% of the reported latencies are within 10.3 ms of the baseline reported value, while taking an order of magnitude less memory.
Unable to display preview. Download preview PDF.