[3437] in Release_7.7_team
Re: NFS rsize and wsize defaults
daemon@ATHENA.MIT.EDU (John Hawkinson)
Tue Aug 27 16:23:35 2002
Date: Tue, 27 Aug 2002 16:23:31 -0400 (EDT)
Message-Id: <200208272023.QAA29595@multics.mit.edu>
To: Bill Cattey <wdc@mit.edu>
CC: release-team@mit.edu
In-reply-to: "[3436] in Release_7.7_team"
From: John Hawkinson <jhawk@MIT.EDU>
| In the early days of attach, MIT Net had routers and gateways with an
| MTU (Maximum Transmission Unit or packet size) of 1500 bytes. To
| prevent lossage from buggy fragmentation/reassembly code, attach was set
| to crowbar NFS filesystems to rsize/wsize of 1024 so that each NFS UDP
| datagram would fit in a single MTU through the gateways.
It is standard-practice to avoid using NFS blocksizes larger than the
MTU in routed networks; this is *not* because of buggy
fragementation/reassembly code, but because a single packet-drop is a
lot more likely across a routed network, as is reordering, and that
causes the entire block to be dropped (possibly after a timeout),
which is very painful.
This is quite a current concern, especially given that the Athena
environment covers many networks of extremely poor quality and high
loss rates (eg. dormitory networks).
| I propose that we remove that default and let rsize and wsize be
| unspecified so that the OS running NFS can use its own best judgement
| about rsize and wsize.
I think this would be a rather poor idea.
Before proposing this seriously, could you please do some tests to see
what sizes are chosen by the modern Athena operating systems by
default, and confirm that they do not fail catastrophically in the
face of steady-state 5% packet loss, a frequent occurance on today's
MITnet?
--jhawk
