Hi, > -----Original Message----- > From: Joe Touch [mailto:[email protected]] > Sent: Thursday, April 30, 2015 7:22 AM > To: Xuxiaohu; Templin, Fred L; Lucy yong; Tom Herbert > Cc: [email protected] > Subject: Re: [Int-area] Why combine IP-in-UDP with GUE? > > > > On 4/29/2015 8:01 PM, Xuxiaohu wrote: > > Hi Templin, > ... > >> Existing tunnel protocols (IP*-in-IP*) are deficient in not providing a > >> tunnel > >> fragmentation mechanism per Section 3.1.7 of RFC2764. > > > > You may have noticed a fact that most modern routers and switches > > have been capable of processing jumbo frames for a long while. In > > other words, fragmentation could be completed avoided in most modern > > networks. > > Jumbo frames are just a larger maximum frame size. Let's say that size is N. > > We're talking about IP-in-IP, IP-in-UDP-in-IP, IP-in-GRE-in-IP, etc. In > all cases, you have: > > N inside N+header > > Unless the headers are 0 bytes long, the encapsulated result is always > larger than the interior contents. > > I.e., if N is your limit, you've now exceeded it. The consequence is that: > > any protocol that has non-zero headers and a packet maximum > always needs fragmentation to be encapsulated in itself > > > Even in the sparse network environments where fragmentation > > is still unavoidable, the default configuration which have been > > widely supported by most vendors (see > > https://tools.ietf.org/html/draft-ietf-intarea-gre-mtu-03#page-4) > > should be enough in most cases. That's the reason why that's > > implemented by many vendors as the DEFAULT configuration, IMHO. > > It's the default because it maximizes router vendor profit and/or > reduced router vendor cost, not because it maximizes network flexibility > or capability. > > > BTW, > > it's preferable to avoid reassembling fragments at the tunnel egress > > due to the negative impact on the forwarding performance, AFAIK. > > It is inherently unavoidable for IPv4 packets with DF=1 or all IPv6 packets. > > > As > > such, it's not recommended to perform fragmentation on the tunnel > > layer and the outer IP layer. > > It's provably required for IPv4 DF=1 and IPv6. > > That's why we're trying to address it.
Joe is right. True that larger MTUs are being deployed especially in the core, but there are still only two magic numbers: 1280 and 1500. Tunnels that can't do 1280 can be declared non-operational. Tunnels that can't do 1500 when there may be PMTUD black holes need fragmentation. Thanks - Fred [email protected] > Joe _______________________________________________ Int-area mailing list [email protected] https://www.ietf.org/mailman/listinfo/int-area
