On 2014-04-16 17:34 , Chris Welti wrote:
Am 16/04/14 13:08, schrieb Jeroen Massar:
On 2014-04-16 11:05 , Chris Welti wrote:
Hi Jeroen,
I can't follow what you're saying. As long as the announcements for the /18 and the /19s are valid, there is no reason for broken connectivity.
Of course there is. The /19 has precedence. Thus for sites that receive the /19 they will route over a much longer route than they should.
I'm aware of that, however that doesn't mean the path is broken, it's just different :) You said the connectivity was broken, not that the chosen paths were not optimal.
Pedantic, but okay, it is indeed not "broken" but definitely not "optimal".
But because it is not optimal it causes broken connectivity if the latency is 300ms more than it should be...
The reason for this is simple and happens all the time.
You mean like Swisscom who internally aggregate all prefixes they receive? :)
Some ISPs will filter more-specific announcements for policy reasons, some won't update their prefix-filter lists regularly, someone makes a mistake and so on. So if an ISP on the "optimal" path filters the /19 you will fall back to the next best path for the /19. that is not filtered. That doesn't mean something is wrong per se.
And in this case the more specific /19 is the 'non-optimal' path, and the /18 is the correct one.
Thus the ISPs that filter these more-specifics are 'winning' while the ones that accept the path are having issues as they are going over the non-optimal path.
Note again, that the most-specific (unless config tweaked) wins.
And the more-specific, is the issue here...
[..]
As shown here:
That doesn't show anything about end-to-end connectivity, just the AS paths.
Which clearly shows that various ISPs around the world are receiving the more-specific and thus the non-optimal path...
All the sites that do receive the /19, especially for instance the 40% of the networks that receive it over for instance Sao Paolo, Brazil, or Moscow IX and do not get it more local will take the long trip around.
Nobody forces them to accept the more specifics... ;)
But they are being asked to do so; and if you are nice you do and you lose out.
Which is affecting connectivity for various folks.
Some of the instances are getting the same path for /18 and /19 and some have longer AS path, but that doesn't necessarily mean the path is worse for all of them.
Not all indeed, but there are cases where it is worse.
[..]
More specifics are evil and give weird routing in various locations.
Don't get me wrong, I'm a fan of getting rid of more-specifics so the global routing table is not polluted.
The "more specifics are evil" is not even because of routing pollution, it is all about the fact that some ISPs do filter and others do not and thus the routing will be indeterminate, one just have to have luck if the path goes well or not.
Greets, Jeroen