Subject: Re: NFS writes NetBSD vs FreeBSD
To: None <port-alpha@netbsd.org>
From: Stephen Jones <smj@cirr.com>
List: port-alpha
Date: 07/15/2004 21:17:04
> On Thu, Jul 15, 2004 at 06:03:49PM -0700, Jason Thorpe wrote:
>> No, it's actually the stupid 82557. It requires a atomic write to a
>> non-aligned region in order for proper operation. Fine on IA, not
>> fine
>> on almost everything else.
Ahh, you guys had pointed that out to me in the past before, but I just
missed it when I used the GENERIC 2.0x kernel .. sorry about that.
Earlier
today I compiled a kernel with -mbwx (as I've always done with 1.6.x
production)
and its doing much better.
I'm also running MP with the little pmap.c patch that Michael Hitch has
been
passing around.
I did double check the interface settings in both SRM and with ifconfig
to confirm
they were in full-duplex mode. I'm not using an etherswitch, but a
cross over cable
between the two.
Here are current messages logged:
Jul 15 23:03:01 vinland /netbsd: nfs_send: proc botch: rep 0x0 arg 0x0
curproc 0xfffffc00053ccee8
Jul 15 23:05:25 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:05:32 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:11:09 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:11:09 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:11:44 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:11:44 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:21:09 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:21:09 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:21:47 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:21:56 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:24:37 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:24:37 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:26:19 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:28:39 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:30:08 vinland /netbsd: nfs server iceland1:/tmp: not
responding
Jul 15 23:30:08 vinland /netbsd: nfs server iceland1:/tmp: is alive
again
Jul 15 23:11:45 iceland kernel: nfs server vinland1:/tmp: not responding
Jul 15 23:11:45 iceland kernel: nfs server vinland1:/tmp: is alive again
(yeh, thats it)
I don't have -D set, so its using the default.
I started the two off at nearly the same moment:
Client Server Low B/s High B/s Avg B/s
Iterations Generated Zeros
FreeBSD NetBSD 2329580 3621499 2816192 203
200MB
NetBSD FreeBSD 1911123 2913115 2324700 168
200MB
So, no more timeouts and it looks like both are getting better writes
in to each other.
Also, directory listings return back a little faster on both .. but I'm
sure we all knew
that it would probably happen once the timeouts were sorted.
FreeBSD netstat -i: (However I don't think these were cleared even
though I went
back to single user mode, turned off networking and came back up).
Name Mtu Network Address Ipkts Ierrs Opkts
Oerrs Coll
fxp0 1500 <Link#1> 00:02:56:00:03:f1 323981216 280944 280391835
0 6115465
fxp0 1500 fe80:1::202:5 fe80:1::202:56ff: 7 - 72
- -
fxp0 1500 10 iceland1 34432002 - 38290660
- -
NetBSD netstat -i:
Name Mtu Network Address Ipkts Ierrs Opkts
Oerrs Colls
fxp0 1500 <Link> 00:02:56:00:0f:ad 42739578 1065 38766502
0 0
fxp0 1500 fe80:: fe80::202:56ff:fe 42739578 1065 38766502
0 0
fxp0 1500 10/8 vinland1 42739578 1065 38766502
0 0
If I let them write on each other one at a time, the FreeBSD writes are
between 5102KB/s
and 5434KB/s .. NetBSD writes are 4225KB/s to 4416KB/s
I'm going to let the NetBSD machine run all night with a fairly heavy
load. The FreeBSD SMP
CS20 has been up for 4 days (I've not seen it crash yet actually) .. I
hope NetBSD can match it.