2004-02-24 17:07:54

by Alex Wun

[permalink] [raw]
Subject: NFS sync write performance

Hi.

I'm interested in the knowing what kind of performance numbers people are getting with NFSv3 synchronous writes.
(i.e. what should I be expecting?)

My system:

RedHat9.0 linux-2.4.20
Dual Intel Xeon P3 1.1Ghz
1GB ram
7-disk Raid5
XFS Filesystem
IntelPRO1000 GigE

Sequential local writes (rewrites) to a 2GB file on the Raid5 go at over 100MB/s. But
over NFS (v3, sync) it's less than 10MB/s. Is this roughly what's expected or is something
wrong with my setup?

As a side note, after modifying some code in the nfs client, nfs server, and xfs filesystem, I was able to
get performance up to ~45MB/s on the 7-disk Raid5 and ~60MB/s on a 7-disk Raid0. These customizations
are still very primitive but they might be worth looking into. I know the nfs developers are busy, but would
they be interested in providing some feedback?

Thanks,
Alex Wun
[email protected]
Consensys Computers Inc.


2004-02-24 20:43:31

by Lever, Charles

[permalink] [raw]
Subject: RE: NFS sync write performance

hi alex-

sync writes are slow, that should be expected. the Linux NFS client
writes a page at a time, for architectural reasons, but could be improved
if we could get it to write a whole "wsize" chunk at a time.


-----Original Message-----
From: Alex Wun [mailto:[email protected]]
Sent: Tuesday, February 24, 2004 9:07 AM
To: [email protected]
Subject: [NFS] NFS sync write performance


Hi.

I'm interested in the knowing what kind of performance numbers people are getting with NFSv3 synchronous writes.
(i.e. what should I be expecting?)

My system:

RedHat9.0 linux-2.4.20
Dual Intel Xeon P3 1.1Ghz
1GB ram
7-disk Raid5
XFS Filesystem
IntelPRO1000 GigE

Sequential local writes (rewrites) to a 2GB file on the Raid5 go at over 100MB/s. But
over NFS (v3, sync) it's less than 10MB/s. Is this roughly what's expected or is something
wrong with my setup?

As a side note, after modifying some code in the nfs client, nfs server, and xfs filesystem, I was able to
get performance up to ~45MB/s on the 7-disk Raid5 and ~60MB/s on a 7-disk Raid0. These customizations
are still very primitive but they might be worth looking into. I know the nfs developers are busy, but would
they be interested in providing some feedback?

Thanks,
Alex Wun
[email protected] <mailto:[email protected]>
Consensys Computers Inc.

2004-02-24 22:07:08

by Tom McNeal

[permalink] [raw]
Subject: RE: NFS sync write performance

I haven't done any measurements in a long time, but I'd think the
differential between sync and async V3 would be minor in general.
Well, with the buffer size limits, it *is* significant, and HPUX definitely
got a big boost when they (well, it was we at the time) went from
8k to 32K buffer sizes, but its nowhere near the differential
between sync and async V2. Has anyone figured out some general numbers?

Regards -

Tom

--
Tom McNeal
(650)906-0761(cell)
(650)964-8459(fax)



-------------------------------------------------------
SF.Net is sponsored by: Speed Start Your Linux Apps Now.
Build and deploy apps & Web services for Linux with
a free DVD software kit from IBM. Click Now!
http://ads.osdn.com/?ad_id=1356&alloc_id=3438&op=click
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

2004-02-24 22:22:55

by Trond Myklebust

[permalink] [raw]
Subject: Re: NFS sync write performance

P? ty , 24/02/2004 klokka 09:07, skreiv Alex Wun:
> As a side note, after modifying some code in the nfs client, nfs server,
and xfs filesystem, I was able to
> get performance up to ~45MB/s on the 7-disk Raid5 and ~60MB/s on a
7-disk Raid0. These customizations
> are still very primitive but they might be worth looking into. I know
the nfs developers are busy, but would
> they be interested in providing some feedback?

Mind explaining what you did to tweak the client/server? If most of the
change was due to setting the "async" export option, then the above
increase would be expected (as is the accompanying reduction in server
failure survival rate).

Cheers,
Trond





-------------------------------------------------------
SF.Net is sponsored by: Speed Start Your Linux Apps Now.
Build and deploy apps & Web services for Linux with
a free DVD software kit from IBM. Click Now!
http://ads.osdn.com/?ad_id=1356&alloc_id=3438&op=click
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs