Hi, Larry, > On Apr 13, 2015, at 9:05 PM, Larry Kreeger (kreeger) <[email protected]> > wrote: > > Hi Carlos, > > Thanks for the reference below. So, this means the answer is "it depends" on > the packetization layer above UDP, but do you know if it is common for > applications running over UDP to pay attention to the PMTU and adjust its > packetization size accordingly? Are there known "problem protocols" that > don't do this? (e.g. RFC 1191 that you reference mentions NFS, but this may > no longer be true of newer implementations?).
That’s right ― applications running over UDP can use the IP-layer info minus their overhead to set their packetization sizes, and/or register to listen to ICMPs with embedded errored packets for them. One example of a protocol doing this, spec’ed in detail, is L2TP / L2TPv3 (e.g., RFC 3931, S4.1.4) Thanks, ― Carlos. > > Thanks, Larry > > From: "Carlos Pignataro (cpignata)" <[email protected] > <mailto:[email protected]>> > Date: Saturday, April 11, 2015 5:55 PM > To: Larry Kreeger <[email protected] <mailto:[email protected]>> > Cc: Anoop Ghanwani <[email protected] <mailto:[email protected]>>, > Erik Nordmark <[email protected] <mailto:[email protected]>>, "[email protected] > <mailto:[email protected]>" <[email protected] <mailto:[email protected]>>, Lucy yong > <[email protected] <mailto:[email protected]>>, Lizhong Jin > <[email protected] <mailto:[email protected]>> > Subject: Re: [nvo3] Encapsulation considerations > > >> On Apr 10, 2015, at 3:49 PM, Larry Kreeger (kreeger) <[email protected] >> <mailto:[email protected]>> wrote: >> >> I thought Path MTU discovery was used to set the MSS in the TCP stack. I >> just wasn't sure if it worked the same way for UDP. >> > > Since, as you know, UDP has no MSS and no connection, the packetization size > state is maintained by the application on top of UDP, or suffer IP-level > fragmentation. > > See e.g, first para of https://tools.ietf.org/html/rfc1191#section-6.1 > <https://tools.ietf.org/html/rfc1191#section-6.1> > > Thanks, > > ― Carlos. > >> - Larry >> >> From: Anoop Ghanwani <[email protected] <mailto:[email protected]>> >> Date: Friday, April 10, 2015 12:10 PM >> To: Larry Kreeger <[email protected] <mailto:[email protected]>> >> Cc: Lizhong Jin <[email protected] <mailto:[email protected]>>, Lucy >> yong <[email protected] <mailto:[email protected]>>, Erik Nordmark >> <[email protected] <mailto:[email protected]>>, "[email protected] >> <mailto:[email protected]>" <[email protected] <mailto:[email protected]>> >> Subject: Re: [nvo3] Encapsulation considerations >> >> Even for TCP it depends on what the MSS is. >> >> Anoop >> >> On Fri, Apr 10, 2015 at 11:55 AM, Larry Kreeger (kreeger) <[email protected] >> <mailto:[email protected]>> wrote: >>> On 4/9/15 7:22 PM, "Lizhong Jin" <[email protected] >>> <mailto:[email protected]>> wrote: >>> >>> >>-----Original Message----- >>> >>From: Lucy yong [mailto:[email protected] >>> >><mailto:[email protected]>] >>> >>Sent: 2015年4月9日 22:28 >>> >>To: Lizhong Jin; 'Erik Nordmark'; [email protected] <mailto:[email protected]> >>> >>Subject: RE: [nvo3] Encapsulation considerations >>> >>Lizhong, >>> >>[snip] >>> >> [Lizhong] If the NVE and tenant is integrated into one device, then >>> >>the issue >>> >>could be solved by implementation. Because tenant know the entropy value >>> >>of >>> >>the first segment, and use the same value to the subsequent segment. So >>> >>different implementation model could provide different entropy value. Or >>> >>do we >>> >>need other mechanism to mitigate this issue, e.g., fragment on NVE in >>> >>draft-herbert-gue-fragmentation. >>> >>[Lucy] IMO: NVO3 solution SHOULD avoid packet fragmentation. >>> >>Draft-herbert-gue-fragmentation provides an option for a GUE application >>> >>to do >>> >>fragmentation but does not require doing it. GUE application decides if >>> >>the >>> >>fragmentation is needed or not. We should not separate two. >>> >[Lizhong] fragmentation could not be avoided, because we are unable to >>> >prevent >>> >the tenant from fragmentation. This is the factor which makes the hashing >>> >based >>> >load balancing unoptimized. >>> >>> I'm not very familiar with host stacks. Do they actually fragment at the >>> IP layer, or is it done at the transport layer before adding the IP >>> header? I'm sure TCP must break the segments up before IP would fragment, >>> but I'm not sure about UDP. >>> >>> - Larry >>> >>> _______________________________________________ >>> nvo3 mailing list >>> [email protected] <mailto:[email protected]> >>> https://www.ietf.org/mailman/listinfo/nvo3 >>> <https://www.ietf.org/mailman/listinfo/nvo3> >> >> _______________________________________________ >> nvo3 mailing list >> [email protected] <mailto:[email protected]> >> https://www.ietf.org/mailman/listinfo/nvo3 >> <https://www.ietf.org/mailman/listinfo/nvo3> >
signature.asc
Description: Message signed with OpenPGP using GPGMail
_______________________________________________ nvo3 mailing list [email protected] https://www.ietf.org/mailman/listinfo/nvo3
