[Cialug] NUMA v. non-NUMA on a processor basis

kristau kristau at gmail.com
Sat Jun 19 11:46:33 CDT 2010


My understanding is that NUMA helps optimize RAM usage by 'preferring' RAM
which is physically closer to the CPU on which the instructions are running.
With a single, multi-core processor, I wouldn't think there's much
advantage. With more than one processor, you often have your RAM slots split
with each proc getting its own bank or bus. If instructions or data are
located on the other processor's bus, there is a time penalty incurred on
every access.

Virtualization technologies can request affinity to a processor to keep a
particular vm on a single processor, but if the amount of RAM needed exeeds
that available locally to the processor, then the non-preffered RAM on the
other processor's bus will be used, resulting in a small performance hit.

That, at least, is my understanding. Please correct me if I'm wrong!

On Jun 18, 2010 2:32 PM, "Nathan C. Smith" <nathan.smith at ipmvs.com> wrote:


Sort of a virtualization question I guess.

Is using NUMA still preferred when doing virtualization on a single
processor machine with multiple cores, or is NUMA purely for multiple
physical processor clusters?

What about NUMA when virtualization is not in use?  Any benefit or
performance penalty on a single or multiple processor machine?

Thanks.

-Nate
_______________________________________________
Cialug mailing list
Cialug at cialug.org
http://cialug.org/mailman/listinfo/cialug
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://cialug.org/pipermail/cialug/attachments/20100619/70ba3c1a/attachment.htm 


More information about the Cialug mailing list