Date: Wed, 15 May 2013 10:47:49 -0400
From: "J. Bruce Fields" <bfields@fieldses.org>
To: James Vanns <james.vanns@framestore.com>
Cc: Linux NFS Mailing List <linux-nfs@vger.kernel.org>
Subject: Re: Where in the server code is fsinfo rtpref calculated?
Message-ID: <20130515144749.GI16811@fieldses.org>
References: <20130515141508.GH16811@fieldses.org>
 <1995711958.19982333.1368628467473.JavaMail.root@framestore.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
In-Reply-To: <1995711958.19982333.1368628467473.JavaMail.root@framestore.com>
Sender: linux-nfs-owner@vger.kernel.org

On Wed, May 15, 2013 at 03:34:27PM +0100, James Vanns wrote:
> 
> > On Wed, May 15, 2013 at 02:42:42PM +0100, James Vanns wrote:
> > > > fs/nfsd/nfssvc.c:nfsd_get_default_maxblksize() is probably a good
> > > > starting point.  Its caller, nfsd_create_serv(), calls
> > > > svc_create_pooled() with the result that's calculated.
> > > 
> > > Hmm. If I've read this section of code correctly, it seems to me
> > > that on most modern NFS servers (using TCP as the transport) the
> > > default
> > > and preferred blocksize negotiated with clients will almost always
> > > be
> > > 1MB - the maximum RPC payload. The nfsd_get_default_maxblksize()
> > > function
> > > seems obsolete for modern 64-bit servers with at least 4G of RAM as
> > > it'll
> > > always prefer this upper bound instead of any value calculated
> > > according to
> > > available RAM.
> > 
> > Well, "obsolete" is an odd way to put it--the code is still expected
> > to work on smaller machines.
> 
> Poor choice of words perhaps. I guess I'm just used to NFS servers being
> pretty hefty pieces of kit and 'small' workstations having a couple of GB
> of RAM too.
> 
> > Arguments welcome about the defaults, thoodd ugh I wonder whether it
> > would be better to be doing this sort of calculation in user space.
> 
> See below.
> 
> > > For what it's worth (not sure if I specified this) I'm running
> > > kernel 2.6.32.
> > > 
> > > Anyway, this file/function appears to set the default *max*
> > > blocksize. I haven't
> > > read all the related code yet, but does the preferred block size
> > > derive
> > > from this maximum too?
> > 
> > See
> > > > For finfo see fs/nfsd/nfs3proc.c:nfsd3_proc_fsinfo, which uses
> > > > svc_max_payload().
> 
> I've just returned from nfsd3_proc_fsinfo() and found what I would
> consider an odd decision - perhaps nothing better was suggested at
> the time. It seems to me that in response to an FSINFO call the reply
> stuffs the max_block_size value in  both the maximum *and* preferred
> block sizes for both read and write. A 1MB block size for a preferred
> default is a little high! If a disk is reading at 33MB/s and we have just
> a single server running 64 knfsd and each READ call is requesting 1MB of
> data then all of a sudden we have an aggregate read speed of ~512k/s

I lost you here.

> and 
> that is without network latencies. And of course we will probably have 100s of
> requests queued behind each knfsd waiting for these 512k reads to finish. All of a
> sudden our user experience is rather poor :(

Note the preferred size is not a minimum--the client isn't forced to do
1MB reads if it really only wants 1 page, for example, if that's what
you mean.

(I haven't actually looked at how typical clients used rt/wtpref.)

--b.

> Perhaps a better suggestion would be to at least expose the maximum and preferred
> block sizes (for both read and write) via a sysctl key so an administrator can set
> it to the underlying block sizes of the file system or physical device?
> 
> Perhaps the defaults should at least be a smaller multiple of the page size or somewhere
> between that and the PDU of the network layer the service is bound too.
> 
> Just my tuppence - and my maths might be flawed ;)
> 
> Jim
> 
> > I'm not sure what the history is behind that logic, though.
> > 
> > --b.
> > 
> 
> -- 
> Jim Vanns
> Senior Software Developer
> Framestore