Single IP routing problems through Level3
peiffer at umn.edu
Sun Jun 15 11:34:31 CDT 2008
Matt Palmer wrote:
> We're seeing some really weird issues with connections that go through / to
> Level3 IP space. Basically, certain "pairs" of IPs (particular L3 IPs
> coupled with particular IPs of ours) have dodgy/nonexistent connectivity,
> but if you change the IP at either end everything's hunky dory.
> I've sniffed (from both ends) pings going from a host in L3 space to our end
> and seen the pings arrive at our end and head back in the direction of L3,
> but they never get to their destination. Traceroutes from L3 stop at the
> next-to-last hop, while traceroutes back get to the hop before L3 space and
> All of this behaviour is source/dest *pair* specific -- if I ping/traceroute
> from another address (in the same netblock as the problematic IP, so all the
> same equipment is involved) at either end, or to another address (again,
> same netblock) at either end, it all works again.
> I've got two questions:
> 1) Has anyone else seen similar behaviour from L3 (or other providers,
> even), so I know I'm not going mad?
> 2) What sort of configuration problem or software bug would cause this sort
> of problem to occur? If it was an IP blacklist (or even a block routing
> issue) anywhere along the line, surely it wouldn't be sensitive to
> changing the other end's address to another one in the same /24?
> Any insight/anecdotes/etc would be greatly appreciated, as it's starting to
> do my head in. Just knowing I'm not alone with this insanity would be nice
> at this point. <grin>
> If it makes any difference, the blocks I'm working from at my end are
> Internap, in 18.104.22.168/23 (we don't have all of it, just most of it),
> while the far end is 22.214.171.124/24.
> - Matt
We commonly see this sort of problem with Layer2 or Layer3 bonded
etherchannel (LACP also). One member of the channel is failing for one
reason or another and dropping traffic. The channel is really not a
load balance mechanism, but is a frame distribution mechanism. The
distribution of frames uses the source and destination IP addresses to
hash out to a particular channel member, and that distribution provides
a rough balance. The problems noted affect traffic in one direction
differently as it is likely assymetric across the channel. Only traffic
across the ailing member will be impacted.
The above can present itself anywhere in the path if channeling is used.
Networking and Telecommunications Services
University of Minnesota/NorthernLights GigaPOP
More information about the NANOG