Curtis Villamizar said the following on 12/29/2006 01:34 PM:
In message <[EMAIL PROTECTED]>
Naiming Shen writes:
It is unfortunate that routers still allow per-packet ECMP to be
configured. At least providers know not to use this feature. But we
can all be sure that somewhere out there someone ignorant of the
problems has enabled it.
I completely agree. Per-packet ECMP should not be used normally,
especially in provider's networks. But I did experience some cases
( a while ago ) on international links, there were some dominant
prefixes which led to very uneven link bandwidth usages due to flow
based algorithm. I think if folks in the routing area still think
things like anycast having this issue, then to signal this attribute is
useful and can be easily done; otherwise, there is absolutely no issue
in terms of anycast prefiexes(ipv4 or ipv6).
Maybe you missed the point. The alternative to per-packet ECMP is not
per-prefix ECMP. The hash based ECMP is only uneven if one or a small
number of host pairs dominate the traffic flow.
The only time that it mattered was when one dominante vendor didn't
offer hash based ECMP. Per-prefix ECMP could be very uneven. That
was a decade ago.
The majority of the router software now is certainly 4-tuple based hash,
but it does not mean in all the cases even load distribution can be
achieved.
As far as anycast goes even with per-packet ECMP, everything is fine
if the application can send one packet that is returned with the
unicast address of an anycast server and use the unicast address from
then on. This sort of probe can be repeated periodically if
Sure, then this involves application awareness of anycast knowledge
and explicit configuration, it's error-prone to say the least.
That would certainly be a big restriction.
necessary. A load influenced anycast could be done within an IGP by
changing a cost but load might tend to slosh (all to one server, then
another). What might work better for that would be an underlying
multicast with limited TTL (reaches multiple close servers, all of
which respond with their loadings). Limited deployment of multicast
might be an issue and non-optimality of a multicast tree with fixed RP
might also be an issue. There are plenty of tools in the tool box to
implement an anycast end user service without resorting to a new
address type and any change to routing.
The WG is OSPF and we're off topic at this point.
Agreed.
- Naiming
Curtis
_______________________________________________
OSPF mailing list
[email protected]
https://www1.ietf.org/mailman/listinfo/ospf
