[linux-cifs-client] Re: [cifs-protocol] Re: CIFS/NFS performance of netapp filer vs local dir

[Matthew Seitz]

Chuck McIntyre wrote:
> On 3/1/07, Matt Seitz <seitz at neopathnetworks.com> wrote:
> Have you seen the post I made to cifs-protocol in the thread entitled
> "Problem - CIFS network throughput vs. NFS"?
> 
> I used bonnie++ from sourceforge:
> http://sourceforge.net/projects/bonnie/ (a disk load benchmarking
> application on CIFS vs. NFS) to generate results from the same client
> against the same server just mounted in two different ways. The
> results are staggeringly bad for CIFS vs. NFS. 

Yes, I did.  Unfortunately, that doesn't answer the question of why there is 
such a performance difference.  It could be the CIFS client, it could be the 
server has better NFS server software than CIFS server software (you don't 
mention what the server software is).  If this is not an isolated network, it 
could be an interaction with other clients or network traffic.

> As you mention, it could be the CIFS client implementation, as
> opposed to something inherent in the protocol, but I am having a hard
> time believing it since it would also have to be something in the
> SMBFS client.

It would be curious what the results would be with a Windows client with the 
same hardware accessing a Windows server.

Of course, even that would not necessarily be proof.  What is really needed is 
an examination of a network trace and client and server logs to determine what 
is causing the performance difference, and then determining whether that was 
something inherent in the protocol, or a limitation of the client or server 
implementation.

> Can you recommend a "good" CIFS client that I can run the same
> benchmark against?

The Microsoft Windows CIFS clients is probably the best.  The problem is trying 
to get a fair apples-to-apples comparison of a benchmark running on two 
different machines with two different operating systems.

For Linux, the standard CIFS clients have been mostly noble one-man operations 
with limited resources and a limited user base.  The "smbfs" client has been all 
but orphaned.  Steve French is doing an admirable job with "cifs_vfs", but I 
don't know how much optimization work he has done.  There is at least  one 
commercial CIFS client for Linux, "Sharity":

http://www.obdev.at/products/sharity

At this point, the question really isn't "Why is CIFS slower than NFS?", it's 
"Why are the Linux CIFS clients (smbfs, cifs_vfs) slower than the Linux NFS 
client?".  Given that, it might be more appropriate to continue this discussion 
on the Linux cifs_vfs mailing list, where the developers can give a more 
informed answer.

-- 
Sincerely,
Matthew Seitz
Customer Support Manager
NeoPath Networks, Inc.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: seitz.vcf
Type: text/x-vcard
Size: 306 bytes
Desc: not available
Url : http://lists.samba.org/archive/linux-cifs-client/attachments/20070301/93d92ff2/seitz.vcf