On Mon, Jul 22, 2002 at 08:03:52AM -0400, Steven M. Bellovin wrote:
> In message <20020720013228.A8641@pyy.jmp.fi>, Jukka Marin writes:
> There have been a number of problems caused by the new default NFS 
> block size.  Should that be backed out?  Better documented, with a 
> work-around?  Made into a sysctl-settable value?

The larger default size is present in 1.5.2 and 1.5.3 as well, and
is only used on i386. Very few problems were reported with 1.5.2,
except for some old ethernet cards, which was more or less expected.
The fix for this is documented in the mount_nfs manual page.

A sysctl would be overkill, since read and writesize is already
a parameter to mount_nfs.

More recently, there seems to be a stream of reports from 1.6BETA
systems, which somehow makes me think that UBC has a different write
pattern (more bursty perhaps), which makes more network hardware
get into trouble. In fact, problems have been reported even with
8k packet size, which had been the old default writesize forever.

I also saw one claim that a 32k packet size was used on sparc,
which can't happen, since sparc still defaults to 8k, as before,
so this is somewhat puzzling (the report is either wrong, or
a recent change somehow violates the write size constraint).

I think that I should probably just make TCP the default, it will
have a better behavior for lost packets. On Solaris, TCP is in
fact the default (by virtue of being in /etc/netconfig before
UDP). However, I'd like find out first why all of a sudden
those reports started showing up now, instead of with 1.5.2,
which already had the bumped packet size for i386.

- Frank

-- 
Frank van der Linden                                    fvdl@wasabisystems.com
==============================================================================
Quality NetBSD Development, Support & Service.   http://www.wasabisystems.com/