Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 28 Aug 2003 02:30:25 -0700
From:      Terry Lambert <tlambert2@mindspring.com>
To:        Jason Stone <freebsd-current@dfmm.org>
Cc:        freebsd-current@freebsd.org
Subject:   Re: nfs tranfers hang in state getblck or nfsread
Message-ID:  <3F4DCBB1.2BDA25AC@mindspring.com>
References:  <Pine.NEB.3.96L.1030827090202.23748B-100000@fledge.watson.org> <20030827211535.N3417@walter>

next in thread | previous in thread | raw e-mail | index | archive | help
Jason Stone wrote:
> I'm also seeing a similar problem - I have a cluster of high-volume
> mailservers delivering mail over nfs to maildirs on a netapp.  The cluster
> was all 4-stable, but I decided to mix a couple of 5.1 boxes in to see how
> they would do.
> 
> The 5.1 boxes accepted and queued mail as well as the 4-stable boxes, but
> delivering the mail into the maildirs over nfs, I kept seeing those
> short-lived hangs, and so the queues started to back up as the boxes were
> accepting mail faster than they could deliver it.
> 
> My mounts are all nfsv3 over udp.

UDP has problems, if you lose any packets at all.  The problem is
that the packet reassembly buffer stays full until you retry, and
the retry is out of band, for packets larger than the MTU size.

What happens when you drop the read and write size low enough that
the data and headers fit in a single UDP packet (e.g. according to
"tcpdump")?  Does it "suddenly" become more reliable?

-- Terry



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3F4DCBB1.2BDA25AC>