Server product kernel requirements

[Prarit Bhargava]

On 11/06/2013 11:21 AM, Peter Robinson wrote:
>>> On 10/30/2013 07:50 PM, David Strauss wrote:
>>>> I was, indeed, drawing an arbitrary line, but we must draw the line somewhere.
>>>> Maybe Fedora 23+ have it set far higher. It's easy to adapt over time to support
>>>> the high end of commodity servers while still being desktop-friendly; we don't
>>>> have a long support window.
>>>>
>>>
>>> Fair enough, but my question is, then, why 512?  If it is completely arbitrary
>>> why not jump it to a high number that people have requested before and be done
>>> with it?  Even 1024 would be acceptable to the HPC users I've talked with FWIW.
>>>
>>> Josh, would you be okay with 1024?
>>
>> Maybe?  That seems like it would be fairly reasonable, but knowing
>> what the overhead numbers are would help.  To be clear, right now we
>> have things set thusly for NR_CPUS:
>>
>> arm=8
>> ppc32=4
>> ppc64/ppc64p7=1024
>> s390x=64
>> i686=32
>> x86_64=128
>>
>> I believe our specific discussion here is about x86_64.  I don't think
>> we're going to change i686 to anything higher than what it's set at
>> right now.
> 
> Even with x86_64 I'm not sure most individual HPC devices have that
> many cores from my experience. Most tend to go for most bang for your
> buck and have more nodes. In the enterprise server space the
> cores/threads tend to top out at around 160 at the moment (8 sockets,
> 10 cores, 2 threads). The exception here tends to be the few massive
> NUMA box shippers and I'm not sure many of those would be looking at
> Fedora, and would likely need custom kernels anyway.

The request for 1024 came from an HPC NUMA box company.  HPC users *want* to
test the Fedora environment on their systems.

P.
> 
> Peter