Date:        Fri, 8 Aug 2003 21:45:55 +0200
    From:        Manuel Bouyer <bouyer@antioche.eu.org>
    Message-ID:  <20030808194555.GA1018@antioche.eu.org>

  | On Thu, Aug 07, 2003 at 04:51:01PM +0700, Robert Elz wrote:
  | > There's no reason I can think of why NFS (in particular, as opposed to the
  | > traffic patterns it generates)

  | It can, because NFS over UDP will send busrt of data,

That's what I meant by "traffic patterns it generates"...

  | If a hardware on the path can't handle the burst (not
  | enouth buffers), the mount hang as NFS will send the same burst over and
  | over.

Yes, which is what I think I used to see with rsync & cvsup (and this was
cvsup from a client on wireless in Thailand to the server in Finland, so...)
(rsync was local).

The wi driver fixes removed both of those problems (for me anyway).

I did a quick test - sent large pings from a system with a fairly high
performance 100Mbps ethernet to the wi device (via an AP on the 100Mb
LAN of course).

I got no loss at all with anything up to 19 fragments generated.
Go to the 20th fragment, and from that point on, no answers at all.

This means (I believe) that the wi driver (at least with Lucent silver
cards, which is all I have available to test with easily) can handle
19 back to back fragments.

It doesn't mean that that is its limit - it is just as likely (perhaps
more so) that the relatively cheap AP is dropping packets when they arrive
at it faster than that (after all, it has to buffer the incoming 100Mbps
packets and eventually transmit them at 11Mbps - there has to be some
queue limit there).   [And no, I didn't bother doing the arithmetic, even
if I knew enough about 802.11 transmit procedures, which I don't, to figure
out how many of the 19 packets should have departed the AP already via the
wireless interface before the last of them arrived via the wired interface].

The "discarded fragments" counter at the wi end started increasing when
the pings stopped being answered, so I am pretty sure it was the ICMP echo
packets being lost, rather than the returning ICMP echo response packets.

The wired and wireless nets would have been (for all practical purposes)
quiet, other than the ping packets, while I was doing this test.

kre