Network Statistics Discovery
It is extremely difficult to determine basic router and link properties such as link bandwidth, propagation delay, queuing delay, processing delay, etc. without SNMP. Even with SNMP, it may be difficult to determine some of these values. This document describes how to gather this information using simple ping times.
Work has already been done by Van Jacobson using ping times to determine link bandwidth and propagation delay. Pathchar was the result of this work. However, pathchar does not take into account important variables such as processing delay and queuing delay. It also does not take into account that ICMP messages to a router often take a different path than data packets, and bandwidth estimates are often estimates of the link for the ICMP packets.
Total delay = propagation delay + transmission delay + queuing delay + processing delay.
Propagation delay + transmission delay is constant, so we can take the minimum of the delays and estimate that to be propagation delay + transmission delay. Processing delay is fairly periodic due to the periodic nature of routing updates. If we remove the periodic factor in the delays, we can isolate the queuing delay, which is extremely important when monitoring traffic on a router.
This is a graph of raw ping times taken from at-gw1.ith-1-0-t3.nysernet.net. The y-axis is ping times in msecs, and the x-axis is time in seconds. Here is the same graph with periodic spikes removed using FFT and thresholding techniques:
Some routers do not have as pronounced spikes as the nysernet router, probably because they have a routing processor per port, and thus reduce the processing load. Thus we have isolated queuing delay.
Work in progress
Need to find a way to automate this process. Currently a lot of human eyeballing and tweaking is involved in removing periodicity.
Author: Rachit Siamwalla
This work is part of the new network management research in the CNRG Research Group with Professor S. Keshav.
This page is last modified 10/10/05 02:18:42 PM