Date: Wed, 24 Jul 2002 14:24:33 -0700 (PDT) From: Matthew Dillon <dillon@apollo.backplane.com> To: "Jaime Bozza" <jbozza@thinkburst.com> Cc: <stable@FreeBSD.ORG>, "'Olaf R'" <olaf@keghouse.net> Subject: Re: RE: RE: Abominable NFSv3 read performance / FreeBSD server / Solaris client Message-ID: <200207242124.g6OLOXaM029616@apollo.backplane.com> References: <025c01c23353$f75cc730$6401010a@bozza>
next in thread | previous in thread | raw e-mail | index | archive | help
:
:Ok, I fiddled with the tcp_recv_hiwat parameter on Solaris 8, along
:with the tcp.sendspace parameter on FreeBSD. No matter what, anything
:above 32767 would continually get the delays.
:
:
:Unfortunately, I don't think I can change them with a running system.
:(Have to set them in /etc/system and reboot to make the changes)
:
:But regardless, both are set to 32768. I wouldn't be surprised if
:rsize=32768
:would work fine if I increased both of these values. Here are the
:beginnings
:of a couple of tcpdumps.
:
:FreeBSD to FreeBSD:
:14:52:36.435205 freebsd-client.1700858811 > freebsd-server.nfs: 132
:lookup [|nfs]
:14:52:36.435334 freebsd-server.nfs > freebsd-client.1700858811: reply ok
:...
Umm. Jaime, these tcp dumps show a UDP NFS mount, not a TCP NFS mount.
The first set of TCP dumps showed a TCP NFS mount.
:...
:that the first
:couple of lines are looking up the file. After that, I'm at a loss with
:the Solaris
:dump, other than the fact that I'm pretty sure the "reply ERR" are
:causing delays
:somewhere. (The FreeBSD one is so clean!)
:
:Does FreeBSD attempt to send packets out using a size of NFS_MAXPACKET
:(33172?) all
:the time? Could the "win 33176" be centered around that? From what I'm
:seeing
A UDP NFS mount will send packets that are approximately rsize
or wsize bytes for full block data reads or writes, depending on
the direction. A TCP NFS mount will not send packets larger then
the MTU (i.e. there will be no fragments).
:(correct me if I'm wrong), FreeBSD is sending 8328 bytes in each TCP
:packet when
:Sending to a FreeBSD client.
This most recent TCP dump you made was a UDP mount, not a TCP mount.
:figure
:out what's really going on here. Any pointers on where to research next
:would be
:helpful. I'm pretty sure that it would have to be something on Solaris
:that's
:causing the problem, but I'm at a loss as to what.
:
:Jaime
Well, I really don't think the rsize or wsize really has anything to
do with it. I still feel that the problem is a window-sizing issue
with TCP and the rsize/wsize changes
I will also note that window sizing is irrelevant with a UDP NFS
mount. It is only relevant on a TCP NFS mount. You should be
getting nearly full speed over a UDP mount (using a rsize and wsize
of 8K). If you aren't then there's likely a network problem
somewhere.
What I really need is a tcpdump from the FreeBSD side showing a few
seconds worth of a large transfer for a TCP NFS mount from the
FreeBSD box to a Solaris client. On the FreeBSD it would be something
like:
tcpdump -s 4096 -n -i <interface> -l > dumpfile
-Matt
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-stable" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200207242124.g6OLOXaM029616>