Common operational misconceptions
mohta at necom830.hpcl.titech.ac.jp
Thu Feb 16 00:11:32 CST 2012
Mark Andrews wrote:
> Well you need to go out of your way to get a ICMP PTB for IPv6
> multicast as the default is to fragment multicast packets at the
> source at network minimum mtu (RFC3542 - May 2003). That's not to
> say it won't happen.
Yes, it will happen, because RFC3542 was, as was discussed
in IETF, written not to prohibit multicast PMTUD.
So, the problem is real.
> As for generation of PTB you rate limit them the way you do for
A problem is that a lot of ICMP packet too big against unicast
is generated, because PMTUD requires hosts periodically try to
send a packet a little larger than the current PMTU.
BTW, that's why IPv6, which inhibit fragmentation by routers,
is no better than IPv4 with fragmentation enabled, because,
periodic generation of ICMP packet too big by routers is as
painful as periodic fragmentation by routers.
>> Note also that some network processors can't efficiently
>> distinguish ICMP packets generated against multicast and
> And why do you need to distingish them?
We don't need to. Instead, we can just give up to use PMTUD
entirely and just send packets of 1280B or less. A problem
is that a tunnel over 1280B PMTU must always fragment 1280B
> You look at the inner
> packet not the ICMP source if you want to rate limit return traffic.
That is a possible problem.
Destination address of inner packet is located far inside
of the ICMP (beyond 64B) that it can not be used for
intrinsic filtering capability of some network processors.
More information about the NANOG