Hi all,
I'm looking for hints on how to improve the performance of Linux clients
(RHEL 3.0, 2.4.21-15.EL, 100Mb NIC) talking to a Solaris server (ES250,
SunOS 5.9 Generic_112233-11 sun4u sparc, 100Mb NIC). What I currently
get from my testing (dd if=/dev/zero of=zero.txt bs=2048 count=1000000)
is between 2 and 2.5 MB/s. I upped the rsize/wsize to 32768 (according
to http://nfs.sourceforge.net/nfs-howto/interop.html#SOLARIS) which did
not help, the result was actually worse than not specifying any mount
options at all. Running the dd test locally on the Sun returns a
throughput of 14MB/s (SW RAID 5 on A1000). Iperf gives a decent 85Mb/s
throughput between the server and a client, I think it's safe to say
that the network layer is ok. I did the test using -o
tcp,nfsvers=3,rsize=32768,wsize=32768, too, it made no difference at all.
I hope to at least double the performance I currently measure. Does that
mean I have to get rid of the Sun...?
Any pointers are welcome, thanks.
Greetings,
Marc
-------------------------------------------------------
This SF.Net email is sponsored by the new InstallShield X.
P=E5 on , 02/06/2004 klokka 07:38, skreiv Marc Schmitt:
> that the network layer is ok. I did the test using -o=20
> tcp,nfsvers=3D3,rsize=3D32768,wsize=3D32768, too, it made no difference a=
t all.
Exactly how did you do this test? Could you check that the
tcp,rsize=3D32k,wsize=3D32k also turn up in /proc/mounts, and then retest?
The reason I ask is that there are 2 things that could corrupt your
test:
You used "-oremount" to change from udp to tcp. That won't work!
You may have had the same filesystem mounted as udp on another
mountpoint. In that case the udp filesystem will be replicated on the
new mountpoint.
Cheers,
Trond
-------------------------------------------------------
This SF.Net email is sponsored by the new InstallShield X.
Trond Myklebust wrote:
> P? on , 02/06/2004 klokka 07:38, skreiv Marc Schmitt:
>
>
>>that the network layer is ok. I did the test using -o
>>tcp,nfsvers=3,rsize=32768,wsize=32768, too, it made no difference at all.
>
>
> Exactly how did you do this test? Could you check that the
> tcp,rsize=32k,wsize=32k also turn up in /proc/mounts, and then retest?
>
> The reason I ask is that there are 2 things that could corrupt your
> test:
>
> You used "-oremount" to change from udp to tcp. That won't work!
>
> You may have had the same filesystem mounted as udp on another
> mountpoint. In that case the udp filesystem will be replicated on the
> new mountpoint.
Indeed, the second case could have happened. I ran the test again,
making sure that the export is only mounted once and that the mount
options I specified appear in /proc/mounts. The result is the same, I
get a throughput of 2.44 MB/s. Just to make sure that the client is not
buggy, I made the same test mounting a RH 7.3 server, RHEL 3 default
mount options, it completes in 3'20" which means whooping 10 MB/s.
Marc
-------------------------------------------------------
This SF.Net email is sponsored by the new InstallShield X.