AS-set diffusion experiments 2005-06-20
Last update: Fri 24 Jun 2005
Purpose
The announcements were designed to discover how far large AS-sets are
propagated, and thus how effective our active probing techniques can be, in
today's Internet with today's IPv4 operational practices.
Announcement schedule
The announcements were sent out by all the RIS
route collectors on
2005-06-20 according to the following schedule:
| 84.205.73.0/24 | 84.205.89.0/24 |
10:00 UTC | 12654 12654 {2121} | 12654 12654 {2121} |
10:30 UTC | — | — |
12:00 UTC | 12654 12654 {5*2121} | 12654 12654 {10*2121} |
12:30 UTC | — | — |
14:00 UTC | 12654 12654 {25*1221}* | 12654 12654 {50*1221}* |
14:30 UTC | — | — |
16:00 UTC | 12654 12654 {75*2121} | 12654 12654 {100*2121} |
16:30 UTC | — | — |
18:00 UTC | 12654 12654 | 12654 12654 |
* The announcement of AS-sets containing AS1221 instead of AS2121 was due to a configuration error.
Preliminary Results
The RIS route collectors were queried 29 minutes after each announcement (1
minute before it was withdrawn) to determine which RIS peers received the
announcement and which did not.
Main results
The results show that the length of AS-sets announced does not significantly
influence the acceptance of the prefix, which leads us to believe that
filtering on the length of an AS-set is not a common practice.
There was no significant difference between the 75-AS set and the 100-AS set,
which suggests that Cisco routers do not filter 75-AS sets by default (the
documentation
of the bgp maxas-limit command is ambiguous on the matter).
In more detail
We grouped the RIS peers according to their behaviour:
- Green peers:
- Accepted all announcements, including AS-sets of all lengths.
- Orange peers:
- Accepted only announcements without AS-sets.
This is probably due to filtering on the origin AS.
- Yellow peers:
- Accepted only AS-sets.
This might be due to imperfect origin AS filtering on the part of certain ASes.
- Blue peers:
- Accepted only announcements without AS-sets or with the longest
AS-sets. We are currently investigating the causes of this behaviour; an
obvious (but doubtful) possibility is that these ASes removed a filter on our
test prefix between 14:30 and 16:00 UTC.
- White peers:
- Fall into no other group.
As can be seen, the results are rather noisy, and we are considering repeating
the announcements under different conditions (e.g. announcing the prefix for
a longer period to eliminate the effects of dampening, announcing it from only
one location to eliminate the effects of anycast, etc. etc.) to obtain more
reliable results.