🚀 go-pugleaf

In article <1992Oct7.150707.29127@awdprime.austin.ibm.com>, will@vulcan.austin.ibm.com (Will Fiveash) writes:
> In article <qncmi00@rhyolite.wpd.sgi.com> vjs@rhyolite.wpd.sgi.com (Vernon Schryver) writes:
> >The only IP fragmenting that an FDDI-ether bridge will usually need to
> >do is of UDP packets.  UDP packets are datagrams, and have no sequence
> >numbers of any kind.  If you don't have FDDI-ether-FDDI networks, the
> >initial MSS negotiation in TCP will pick 1500.  This is consistent with
> >the TCP/IP lore (and RFC) that considers fragmenting harmful.
> 
> I am somewhat confused about the statement concerning MSS negotiation.
> My understanding was that if the two connecting systems were on the
> same physical network then the MSS was the MTU of the interface -
> TCP/IP header length.  If they were not on the same physical net, i.e.
> router(s) involved, then the MSS was 536.  This is what I am reading
> from Doug Comer's book "Internetworking with TCP/IP".

Are we disagreeing?  I don't think so.

As I understand BSD style code and the standards,
    -each system should use an MSS that is no larger than the MSS
	negotiated with the MSS option.

    -two systems on the same wire using the same IP network number
	should use the MTU appropriate for the common wire.

    -a pair of systems on different (sub)networks should consider
	using a smaller maximum segment size to prevent fragmentation
	in the gateways between their (sub)networks.  The smaller
	number may be as small as the minimum legal IP maximum packet
	size, which is a lot smaller than 512.  (I've forgotten.)

    -the classic "smaller" size is >= 512 and <= 576.  Notice that
	536 is 40 less than 576.  40 is the familiar TCP/IP magic
	number, the sum of the sizes of a TCP header and an IP header
	without options.

    -it does not matter what physical network the two systems are on.
	The 500 MSS has to do with being on the same IP or logical
	network.  Bridges are invisible and (must be) irrelevant,
	unless they connect a host on one medium with one MTU to
	another host on another medium with a different MTU.

If I were not lazy, I'd go look up RFC chapter and verse.  I vaguely
recall that H.R. has relaxed the "should ... 5xx".  Maybe not.

I do think they say that the systems are free to choose a larger size
if they can avoid causing fragmentation.  After all, that's what "MTU
discovery" is all about.  It is also the purpose of the
"allnetsarelocal" and "allsubnetsarelocal" switches in BSD code.

Using ~500 when you could have used 1500 does terrible things to
performance.  It bogs down gateways not only with 3 times as many data
packets at up to the 3 times the data packet rate, it also as much as
triples the number of TCP ACK's flowing in the opposite direction.

Using ~500 when you should have used >=4K would be an abomination.


Vernon Schryver,  vjs@sgi.com