CONFIG_NR_CPUS

[Ruben Kerkhof]

On Sun, May 20, 2012 at 2:32 PM, Gilboa Davara <gilboad at gmail.com> wrote:
> On Sun, May 20, 2012 at 4:07 AM, Josh Boyer <jwboyer at redhat.com> wrote:
>>
>> On Sat, May 19, 2012 at 02:54:33PM +0200, Xose Vazquez Perez wrote:
>> >
>> > >* Tue Jan 17 2012 Dave Jones <davej at redhat.com>
>> > >- Rawhide builds now use MAXSMP on x86.
>> > >- For release builds, set x86-64 to support 64 CPUs.
>> > >  If larger systems become widespread, we can increase in an update.
>> >
>> > _today_
>> >
>> > amd:   4sockets * 16cores = 64
>>
>> Awesome.  Got that covered still.
>>
>> > intel: 4sockets * 10cores * 2threads = 80
>>
>> Which particular CPU/Motherboard combo is that, and how often do we see
>> it in Fedora?
>>
>> I'm not opposed to bumping it up to 128 or something, but I'm curious
>> how many people are actually going to see benefits.
>>
>> josh
>
> At least in my case I did run Fedora 12-16 on 4S and 8S machines to
> test software scalability on (extreme) high-end hardware.
> Though, IMHO anyone that's crazy enough to run Fedora on a high-end
> 4S/8S machine is more than capable of rebuilding the kernel with
> CONFIG_NR_CPUS 256...
>
> However, given the fact that x86_64 machines tend to be far less
> memory constrained than i686 machines, I doubt that raising the limit
> to 128 will cause too many issues. (Isn't NR_CPUS == 512 in el6?)

It's even higher these days:

grep NR_CPUS /boot/config-2.6.32-220.13.1.el6.x86_64
CONFIG_NR_CPUS=4096

I'm curious, has anyone measured what the memory overhead is of
keeping NR_CPUS at 512?
arch/x86/Kconfig says "This is purely to save memory - each supported
CPU adds approximately eight kilobytes to the kernel image." If that's
true, 512 cpus use 4MB, something I'm willing to live with on my 64bit
 servers.

Thanks,

Ruben