Subject: Re: [PATCH] improve the performance of large sequential write NFS
	workloads
From: Steve Rago <sar@nec-labs.com>
To: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>, linux-nfs@vger.kernel.org,
       linux-kernel@vger.kernel.org, Trond.Myklebust@netapp.com,
       Wu Fengguang <fengguang.wu@intel.com>,
       "jens.axboe" <jens.axboe@oracle.com>
In-Reply-To: <20091218220741.GB21131@elte.hu>
References: <1261015420.1947.54.camel@serenity>
	 <1261037877.27920.36.camel@laptop> <1261164799.1947.123.camel@serenity>
	 <20091218194129.GB6153@elte.hu> <1261171211.1947.135.camel@serenity>
	 <20091218220741.GB21131@elte.hu>
Content-Type: text/plain
Date: Fri, 18 Dec 2009 17:46:56 -0500
Message-Id: <1261176416.1947.149.camel@serenity>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2825
Lines: 63


On Fri, 2009-12-18 at 23:07 +0100, Ingo Molnar wrote:
> * Steve Rago <sar@nec-labs.com> wrote:
> 
> > 
> > On Fri, 2009-12-18 at 20:41 +0100, Ingo Molnar wrote:
> > > * Steve Rago <sar@nec-labs.com> wrote:
> > > 
> > > > > Also, I don't think this needs to have a sysctl, it should just work.
> > > > 
> > > > The sysctl is a *good thing* in that it allows the eager writeback behavior 
> > > > to be tuned and shut off if need be.  I can only test the changes on a 
> > > > finite set of systems, so better safe than sorry.
> > > 
> > > This issue has been settled many years ago and that's not what we do in the 
> > > Linux kernel. We prefer patches to core code where we are reasonably sure they 
> > > result in good behavior - and then we fix bugs in the new behavior, if any. 
> > > 
> > > (Otherwise odd sysctls would mushroom quickly and the system would become 
> > > untestable in practice.)
> > > 
> > > 	Ingo
> > 
> > I don't disagree, but "that's not what we do" hardly provides insight into 
> > making the judgment call. [...]
> 
> I gave you an example of the problems that arise, see the last sentence above.
> 
> > [...]  In this case, the variety of combinations of NFS server speed, NFS 
> > client speed, transmission link speed, client memory size, and server memory 
> > size argues for a tunable parameter, because one value probably won't work 
> > well in all combinations.  Making it change dynamically based on these 
> > parameters is more complicated than these circumstances call for, IMHO.
> 
> So having crappy tunables is the reason to introduce even more tunables? I 
> think you just gave a good second example of why we dont want sysctls for 
> features like this.
> 
> 	Ingo

The examples I cited are not tunables.  They are characteristics of the
systems we use.  I can't squeeze more than 1Gb/s out of my gigabit
Ethernet connection; I can't make my 2GHz CPU compute any faster; I am
limited by these components to the performance I can attain.  Writing
software that performs well in all combinations, especially to take
advantage of the myriad of combinations, is difficult at best.  The
tunable introduced in the patch is a compromise to writing a much more
complicated adaptive algorithm that most likely wouldn't have access to
all of the information it needed anyway.

Steve

> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/