Estimating Router ICMP Generation Delays

The results indicate that the answer to this question is “few” which implies that tools such as pathchar and treno will not in practice suffer greatly from slow-path/fast-path differences, as has been a concern.

Abstract

A number of Internet measurement techniques rely on eliciting ICMP replies from routers inside the network [7], [9]. One question regarding the accuracy of these techniques concerns whether routers introduce delays when generating the ICMPs. We present a technique for estimating these delays. The approach is based on sending packets with spoofed source addresses, some with sufficient TTL to travel all the way to the destination (“direct”), others with a limited TTL that will cause them to be transformed at a given hop into an ICMP Time Exceeded (TE) packet. By spoofing the source address of the probes to match the destination, these TEs are routed identically to how the direct probes are routed. The technique thus allows us to factor out effects due to asymmetric routing or conditions on the return path from a given router back to the source. We have implemented the technique in a tool, fsd (fast-path/slow-path discriminator). In this paper, we discuss an analysis of a set of measurements made using fsd on the NIMI measurement infrastructure [13]. Our analyses provide helpful insight into the question of “what proportion of routers in today’s Internet are slow at generating TE replies?” Our results indicate that the answer to this question is “few” which implies that tools such as pathchar and treno will not in practice suffer greatly from slow-path/fast-path differences, as has been a concern.