blade servers (was Re: [Bioclusters] Any experiences with these guys? reeks)

[Goran Ceric]

In my experience, HT helps the most if you run several competing 
applications. It's not that beneficial to a single multithreaded 
application unless the threads are complimentary to each other. It all 
depends on what you want to run. If you run code that's 100% FP, I think 
you better turn HT off. Sometimes you have to try it with and without HT 
to see what works better. And yes, Tim is right, current kernels don't 
schedule smartly to logical CPUs.

Tim Cutts wrote:

>On 04-Nov-03, Ognen Duzlevski wrote:
>  
>
>>Could someone educate me on the importance of turning off hyperthreading?
>>I guess I've been sleeping in the past several months...
>>    
>>
>
>Current Linux kernels don't know how to schedule to it; they see four
>distinct CPUs.  If you run two single threaded jobs on a machine with
>hyperthreading enabled, it's possible that that the scheduler will
>decide to put both on the same physical CPU, and leave the other one
>free.  This will not happen if hyperthreading is disabled, and each job
>will get a physical CPU to itself.
>
>So, if you run multithreaded code exclusively, or predominantly, switch
>hypethreading on.  If you principally run single threaded jobs, switch
>it off, until such time as you're running a kernel with a scheduler
>which understands hyperthreading (I believe there are some patches to do
>this, now)
>
>Tim
>
>  
>