Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss

Linux and AIX NFS performance

289 views
Skip to first unread message

Bernd Dammann

unread,
Dec 9, 1999, 3:00:00 AM12/9/99
to
I did some NFS performance investigation on a real-world example,
moving a 26MB file from/to a NFS-mounted filesystem:

- the hardware:
AIX client/server : RS/6000 model 375, 64 MB RAM, AIX 4.2.1
AIX client (case 2): RS/6000 model 590, 128 MB RAM, AIX 4.2.1
Linux client/server: Pentium II 450MHz, 128 MB RAM, RedHat 5.2

- the network: 10Mbit/s

- the file:
-rw-r--r-- 1 bernd users 26253720 Nov 20 15:43 bigfile

- the results:

# 1. AIX server, Linux client #############################
#
# mount options: rsize=8092, wsize=8092
#
linux:aix_nfs_dir>time mv bigfile /tmp
0.020u 0.910s 0:28.87 3.2% 0+0k 0+0io 81pf+0w
linux:aix_nfs_dir>time mv /tmp/bigfile .
0.000u 1.250s 6:11.79 0.3% 0+0k 0+0io 6491pf+0w

# mount options: rsize=8092, wsize=2048
#
linux:aix_nfs_dir>time mv bigfile /tmp
0.000u 1.790s 0:34.42 5.2% 0+0k 0+0io 89pf+0w
linux:aix_nfs_dir>time mv /tmp/bigfile .
0.030u 2.530s 11:40.96 0.3% 0+0k 0+0io 6500pf+0w

# mount options: rsize=2048, wsize=2048
#
linux:aix_nfs_dir>time mv bigfile /tmp
0.010u 2.440s 0:43.91 5.5% 0+0k 0+0io 81pf+0w
linux:aix_nfs_dir>time mv /tmp/bigfile .
0.010u 2.620s 11:39.39 0.3% 0+0k 0+0io 6500pf+0w

# 2. AIX server, AIX client #############################
#
# mount options: AIX standard nfs_v3 settings
#
aix_client:aix_nfs_dir>time mv bigfile /tmp
0.080u 2.090s 0:27.00 8.0% 1+2k 0+0io 711pf+0w
aix_client:aix_nfs_dir>time mv /tmp/bigfile .
0.050u 1.270s 0:28.15 4.6% 0+2k 0+0io 105pf+0w

# 3. Linux server, AIX client #############################
#
# mount options: AIX standard settings
#
aix_client:linux_nfs_dir>time mv bigfile /tmp
0.150u 3.790s 0:28.16 13.9% 0+3k 0+0io 3001pf+0w
aix_client:linux_nfs_dir>time mv /tmp/bigfile .
0.060u 2.950s 0:29.07 10.3% 1+2k 0+0io 169pf+0w

# 4. Linux server, Linux client
#
linux_client:linux_nfs_dir>time mv bigfile /tmp
0.020u 1.590s 0:42.27 3.8% 0+0k 0+0io 81pf+0w
linux_client:linux_nfs_dir>time mv /tmp/bigfile .
0.000u 1.900s 0:42.89 4.4% 0+0k 0+0io 6491pf+0w

- the conclusion:

There is a NFS problem between Linux clients and AIX servers,
especially when writing to the AIX server.

- the question:

Who is to blame, Linux or AIX?

--
# Bernd Dammann <be...@fki.dtu.dk> | "Why stop now,
# Department of Physical Chemistry | just when I am hating it?"
# The Technical University of Denmark |---------------------------------
# Building 206 | phone: (+45) 45 25 24 81
# DK-2800 Lyngby, Denmark | http://www.fki.dtu.dk/~bernd/

banger

unread,
Dec 9, 1999, 3:00:00 AM12/9/99
to
Bernd Dammann wrote:

> I did some NFS performance investigation on a real-world example,
> moving a 26MB file from/to a NFS-mounted filesystem:
>

> - the conclusion:
>
> There is a NFS problem between Linux clients and AIX servers,
> especially when writing to the AIX server.
>
> - the question:
>
> Who is to blame, Linux or AIX?
>

Noticed the same problem. Doesn't AIX use NFS version 3 and Linux uses
version 2?
That could be the problem. But I am not sure.


Marshall Lucas

unread,
Dec 9, 1999, 3:00:00 AM12/9/99
to
The problem has been solved for me:

I had to use an rsize of 1024 and a wsize of 8192. That seems to work the
best. Writing is not a problem it's wicked fast. Reading for some reason
(read: I haven't had time to investigate) requires a very small buffer for
Linux to be happy. Maybe a result of using NFS 2 instead of 3. Any ideas?

-- Marshall

banger <ban...@no.where.net> wrote in message
news:384FC533...@no.where.net...

Håkan Winbom

unread,
Dec 9, 1999, 3:00:00 AM12/9/99
to
What kind of network do you have ?
TP ethernet is known to have problems with autonegotiate of half/full duplex and
10 vs 100 mbit
Ensure that both your switch and adapter are set hard to the same value.
e.g. never trust autonegotiate!
/HW
PS I had a system where ftp put/get worked just fine but NFS crawled, this was
resolved
once disable autonegotiate in both ends.


Marshall Lucas skrev:

winbom.vcf

Phil Edwards

unread,
Dec 10, 1999, 3:00:00 AM12/10/99
to

You might also want to try, on the AIX box:

nfso -o nfs_use_reserved_ports=1

Note that this won't survive a re-boot, so you will need to add it on
the end of /etc/rc.tcpip if you want to make it (semi-)permanent

--
Phil Edwards
Technical Specialist
==========================================================================
Travellog Systems Phone +44 (0)1444
459016
The Priory, Haywards Heath Fax +44 (0)1444
456655
West Sussex, RH16 3LB
mailto:edwa...@travellog.co.uk
United Kingdom
http://www.travellog.co.uk
==========================================================================

Cameron Schaus

unread,
Dec 20, 1999, 3:00:00 AM12/20/99
to
In comp.unix.aix Bernd Dammann <be...@lipid.fki.dtu.dk> wrote:
: - the conclusion:

: There is a NFS problem between Linux clients and AIX servers,
: especially when writing to the AIX server.

: - the question:

: Who is to blame, Linux or AIX?

We have the same problem with linux clients writing to our aix432 nfs
server. Writes are painfully slow, especially when linking files
during compilation.

I had some success fixing this problem by applying nfsv3 patches to
the linux kernel, patching the mount command and using the new kernel
to mount nfs file systems. Unfortunatly, the nfsv3 patches did not
like talking to our Solaris nfs server, and caused the machine to
hang. I used the patches a few months ago, so they may be better now.

If you can't find the patches, send me email and I will send you the
url.

Cam


---
Cameron Schaus
sch...@cpsc.ucalgary.ca

0 new messages