Open Stack

Le 25/12/2013 02:00, Dnsbed Ops a écrit :
> On 2013-12-24 16:21, sylecn wrote:
>> I'd like to hear some real world experience on this topic. Do you enable
>> hyper-threading in your openstack cluster? Does it matter at all for
>> your use case?
>>
>> My current use case is java based web app and node.js based web app.
>> Both facing internet users with potentially high concurrency. The cpu on
>> nova-compute node is E5-2620x2, 24 cores with HT on.
>
> I have also the same question on the hyper-threading item.
> Do you anybody have the experience on it? 
I have no *real* experiences, but I think we are speaking about CPU 
context switch cost here.
Hyper-threading allow two process to share 1 CPU computation power, with 
a minimal (from Intel says) context switch cost.
Facts seams to demonstrate that two process, on a long average, fighting 
to get resources on the same CPU with hyper-threading enable (thus on 2 
virtual CPU) will get less than 1/2 (don't know the magnitude order ...) 
of its total power computing, mainly because of the context swift cost.

So it really depends of the workload profile of the host. My opinion 
would be to disable hyper threading on hardware subject to very 
intensive CPU computation needs (angry CPU algorithm). And to enable it 
when there are potentially lots of thread that *some times* need CPU 
computing, like VM for example.
If you are not a HPC guru that really knows what he is doing/what he 
needs, the safest way would maybe to let it enabled ;-)

Just my thoughts, could be not completely true.

Cédric