Message-ID: <47A1D03A.3020508@inria.fr>
Date: Thu, 31 Jan 2008 14:42:18 +0100
From: Brice Goglin <Brice.Goglin@inria.fr>
User-Agent: Mozilla-Thunderbird 2.0.0.9 (X11/20080110)
MIME-Version: 1.0
To: Paul Mundt <lethal@linux-sh.org>, Chris Snook <csnook@redhat.com>,
       linux-kernel@vger.kernel.org
Subject: Re: Purpose of numa_node?
References: <47A11ACD.1090400@redhat.com> <20080131074045.GA13788@linux-sh.org>
In-Reply-To: <20080131074045.GA13788@linux-sh.org>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2316
Lines: 44

Paul Mundt wrote:
> On Wed, Jan 30, 2008 at 07:48:13PM -0500, Chris Snook wrote:
>   
>> While pondering ways to optimize I/O and swapping on large NUMA machines, I 
>> noticed that the numa_node field in struct device isn't actually used 
>> anywhere. We just have a couple dozen lines of code to conditionally 
>>  create a sysfs file that will always return -1.  Is anyone even working on 
>> code to actually use this field?  I think it's a good piece of information 
>> to keep track of, so I'm not suggesting we remove it, but I want to make 
>> sure I'm not stepping on toes or duplicating effort if I try to make it 
>> useful.
>>     
> It's manipulated with accessors. If you look at the users of
> dev_to_node()/set_dev_node() you can see where it's being used. It's
> primarily used in allocation paths for node locality, and the existing
> set_dev_node() callsites are places where node locality information
> already exists (ie, which node a given controller sits on). You can see
> this in places like PCI (pcibus_to_node()) and USB, with node allocation
> hints used in places like the dmapool and skb alloc paths.
>
> The in-kernel use looks perfectly sane in that regard, though I'm not
> sure what the point of exporting this as a RO attribute to userspace is.
> Presumably someone has a tool somewhere that cares about this.
>   

I added the numa_node sysfs attribute in the beginning to make it easier 
to bind processes near some devices. So yes I have some user-space tool 
using it. It is much easier to use than the local_cpus field on large 
machines, especially when you use the libnuma interface to bind things, 
since you don't have to translate numa_node from/to cpumasks.

It works fine on regular machines such as dual opterons. However, I 
noticed recently that it was wrong on some quad-opteron machines (see 
http://marc.info/?l=linux-pci&m=119072400008538&w=2) because something 
is not initialized in the right order. But I haven't tested 2.6.24 on 
this hardware yet, and I don't know if things have changed regarding this.

Brice

--
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/