Prefix lengths are relevant to router vendors because of the need to estimate lookup overhead required for a full route table. Each segment of the address space carries a different average lookup cost and sees different average traffic patterns due to characteristics of address allocation and use.
These graphs are based on a 7 minute OC3mon packet trace collected on March 12, 1998 from FIX West, and a route table from the University of Oregon Route Views project[Meyer97] collected on the same day. At this time, the route table included routes from the following peers:
ANS (Cleveland) 188.8.131.52 through AS1673 CERFnet (San Diego) 184.108.40.206 through AS1740 DIGEX (MAE-EAST) 220.127.116.11 through AS2548 ESnet (GA) 18.104.22.168 through AS293 Europe (RIPE) 22.214.171.124 through AS3333 IAGnet (Chicago) 126.96.36.199 through AS1225 IIJ (Japan) 188.8.131.52 through AS2497 JINX (Johannesburg) 184.108.40.206 through AS2905 LINX (London) 220.127.116.11 through AS5459 MCI (San Francisco) 18.104.22.168 through AS3561 PIPEX (London) 22.214.171.124 through AS1849 Sprint (Stockton) 126.96.36.199 through AS1239 vBNS (Hayward) 188.8.131.52 through AS145 Verio (MAE-WEST) 184.108.40.206 through AS2914 blackrose.org (Ann Arbor) 220.127.116.11 through AS234
Note that these results are directly comparable to the CAIDA paper presented at Inet 98[Claffy98a]. The results are quite similar to figures 11a, 11b, and 12 in that paper despite the fact that the data came from a public exchange point rather than inside a private backbone.
The first two graphs show the distribution of packets and bytes seen relative to the length of the network mask of the network that either generated or received the data. The green line in the first graph superimposes the number of routes for each prefix length in our composite routing table. The largest number of routes have a 24 bit mask (corresponding to the old `class C' networks), but the largest amount of traffic is sourced from and destined to networks with a 16 bit route prefix mask (corresponding to the old `class B' networks).
From a 20-minute trace of IP packets from a wide-area exchange point (FIX-West)
This last graph shows that the longer prefixes are net producers of network traffic compared to the shorter prefixes, as indicated by the larger mean packet size and thus greater average payload size. This suggests the presence of more (content) servers on long prefix networks than on short prefix ones.
figure 3: Mean sizes of packets as a function of the prefix lengths of their source and destination addresses. Based on a 20-minute IP packet trace from a wide-area exchange point (FIX-West).
references1. [Claffy98a]. k claffy, G. Miller, and K. Thompson, the nature of the beast: recent traffic measurements from an Internet backbone. 23 April 1998 (INET '98 presentation).
2. [Meyer97], University of Oregon Route Views Project, http://www.routeviews.org. Advanced Network Technology Center, David Meyer (now at Cisco Systems).
Thanks to Hans-Werner Braun and NLANR/MOAT for supplying the packet trace data, and David Meyer for maintaining the Route Views project. Sean McCreary and kc, 10/26/98