[Top] [All Lists]

Re: NFS performance - more info

To: linux-xfs@xxxxxxxxxxx
Subject: Re: NFS performance - more info
From: Thomas Graichen <news-innominate.list.sgi.xfs@xxxxxxxxxxxxx>
Date: 4 Mar 2001 12:37:13 GMT
Distribution: local
Organization: innominate AG, Berlin, Germany
References: <news2mail-97t0uh$5qj$1@xxxxxxxxxxxxxxxxxxxxxx> <200103041103.LAA23941@xxxxxxxxxxxxxxxxxxxxxxxxxxx>
Reply-to: Thomas Graichen <thomas.graichen@xxxxxxxxxxxxxx>
Sender: owner-linux-xfs@xxxxxxxxxxx
User-agent: tin/1.4.4-20000803 ("Vet for the Insane") (UNIX) (Linux/2.4.2-XFS (i686))
Robin Humble <rjh@xxxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:

> Thomas Graichen writes:
>>Robin Humble <rjh@xxxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
>>* did you try kiocluster's (only work with ide or scsi not with md
>>  or lvm) - no idea if it might help -> mount -o kiocluster ...

> No, I didn't try this. I need to use RAID in the end so it's not
> much use to me... :-/

but it might be a good idea to try it anyway to circle the problem
down a bit

>>* the other good lines posted often here (but i don't think they
>>  will help here) -> mount -o logbufs=4,logbsize=32768 ...
>>* creating the filesystem with a bigger journal is also a good idea
>>  for real workload (ok - i think not in this case too - because
>>  here is not much metadata to journal :-)
>>  -> mkfs -t xfs -l size=32768b ...

> yeah - good idea - I could try that, but like you say - I don't know
> why they would help. It's just one file being written.

>>* did you try the same with an up kernel? - maybe some smp locking
>>  thing somethere

> I just tried the 'dd' between my home machines (server is
> Athlon700/ATA66 single processor, single disk, RH6.2, older 2.4.2 cvs
> kernel, and the client is p120 laptop RH6.2) and found much the same
> behaviour. However because I only have a 10Mbit network at home, the
> nfsd loads didn't get high enough to choke the transfers like they do
> with the 100Mbit connections at work. At home, writing to either XFS
> or ext2 with a dd achieved about ~1MB/s which is what you'd expect.
>   time dd if=/dev/zero of=/mnt/xfs/bigFile bs=1M count=256
> takes ~270s or so.

> For the XFS writes, the nfsd load was about 25% total (4 daemons,
> ~5.5% each), and it was only about 3% total for ext2 (<1% per nfsd).
> So I guess this problem only really shows up on 100Mbit or greater
> networks where the traffic to the daemons is 10x higher.

> So yeah, because I see roughly the same nfsd load behaviour on a
> single procesor box, think that SMP isn't much of an issue, but yes I
> should test it. Is there a way to boot an SMP kernel in uni-processor
> mode or should I recompile?

> Your bonnie setup (in the other email) is possibly a bit flawed in
> that it writes only 128M files and the server has 128M of memory, so
> in theory the server doesn't have to write to disk at all. Could you
> maybe try it again with a file size that's more than twice the server
> memory size? That's what bonnie++ tells you to do anyway, I don't
> know about bonnie.
> Are you using NFSv3 on the server/client? nfsstat will let you know.

oh yes - you are right - i took 128 because the client had 64 - but
you are right the server has 128 and that counts here ... it'S v3

one other idea: does the size of the files matter - so do you get
different results for say 25, 250 and 2500mb (always first run to
ignore caching)?


                                                         innominate AG
                                                  the linux architects
tel: +49-30-308806-13   fax: -77             http://www.innominate.com

<Prev in Thread] Current Thread [Next in Thread>