| 1 |
> On Sun, Dec 16, 2012 at 10:39 AM, J. Roeleveld <joost@××××××××.org> wrote: |
| 2 |
>> On Sunday, December 16, 2012 01:52:46 PM Volker Armin Hemmann wrote: |
| 3 |
>>> Am Samstag, 15. Dezember 2012, 20:57:24 schrieb J. Roeleveld: |
| 4 |
>>> > Even on a system with only 2 sockets, it can be useful to have NUMA |
| 5 |
>>> > available. |
| 6 |
>>> |
| 7 |
>>> or not, because it costs you performance. |
| 8 |
>> |
| 9 |
>> When does it cost performance? |
| 10 |
>> In all situations? |
| 11 |
> |
| 12 |
> It adds some additional logic to memory allocation (put an allocation |
| 13 |
> near the process that uses it) and to process scheduling (keep the |
| 14 |
> process near its memory, but bump it to a more distant idle core if |
| 15 |
> necessary). |
| 16 |
|
| 17 |
That's the way it's supposed to work, yes :) |
| 18 |
|
| 19 |
> In all honestly, it's not a performance loss you're likely to notice, |
| 20 |
> unless you're so in need of squeezing out every spare cycle that you |
| 21 |
> most definitely _have_ hardware where there are disconnected memory |
| 22 |
> banks. I'm not convinced it's even measurable for us mundanes and our |
| 23 |
> hardware. |
| 24 |
|
| 25 |
I don't think I would notice it either, but as the system I have supports |
| 26 |
it, I want to use it. |
| 27 |
And then I want to be certain it actually supports it correctly. |
| 28 |
|
| 29 |
The system I'm talking about is used for testing purposes. Running |
| 30 |
multiple VMs. As far as I know, Xen has support for it, just need to |
| 31 |
configure it properly. |
| 32 |
And for this usecase, I think NUMA with only 2 physical CPUs should make a |
| 33 |
positive difference. |
| 34 |
|
| 35 |
-- |
| 36 |
Joost |
Archives