[prev in list] [next in list] [prev in thread] [next in thread] 

List:       npaci-rocks-discussion
Subject:    Re: [Rocks-Discuss] General question about APPRO equipment and Rocks
From:       Joseph Norris <jnorris () ucmerced ! edu>
Date:       2008-09-30 23:46:41
Message-ID: 48E2BA61.1070805 () ucmerced ! edu
[Download RAW message or body]

Thank you for this explanation.  This helps me understand a bit more.

Philip Papadopoulos wrote:
> On Tue, Sep 30, 2008 at 3:56 PM, Joseph Norris <jnorris@ucmerced.edu> wrote:
>
>   
>> All of these boxes are four processors per box.  Could you elaborate on the
>> last statement?
>>
>> The difference might matter for an application spanning the nodes.
>>
>> What could be the possible applications that might have an issues with a
>> mix of dual and quad core boxes?
>>     
>
> It's all about the other resources being constrained
>
> A dual socket quad-core has 8 CPUs, 8 L1, 8 (or 4) L2 Caches, N (say four)
> memory channels, 1 network connection.
>
> a dual socket dual-core has 4 CPUs, 4 L1 Caches, 4 L2 caches, N (say four)
> memory
> channels, 1 network connection.
>
> The dual core system has twice the available network BW and memory BW - per
> core -
> than the four core system. If your code is primarily memory BW bound, then
> the extra cores get you very little.  The answer of "it depends" is
> operative here.  It depends on your code as to whether more cores is a help
> or a hindrance.   The code will work on
> either system, you just may not get the performance you expect.   Parallel
> codes tend to run relative to their slowest bottleneck, whether it be
> network bw, network latency, memory BW, or CPU speed.
>
>
>
>   
>> thanks.
>>
>>
>>
>>
>> Philip Papadopoulos wrote:
>>
>>     
>>> On Tue, Sep 30, 2008 at 9:40 AM, Joseph Norris <jnorris@ucmerced.edu>
>>> wrote:
>>>
>>>
>>>
>>>       
>>>> Hello group,
>>>>
>>>> Here at UC Merced we have a 67 node cluster that was built and installed
>>>> by
>>>> APPRO well before I came to work as sys admin.  I have now built a
>>>> smaller
>>>> cluster using lower cost servers.  My general question is:  Has anyone on
>>>> the list had any experience with extending APPRO clusters using other
>>>> types
>>>> of servers as compute nodes?
>>>>
>>>>
>>>>         
>>> Depends on what Cluster Mgmt you are running on your Appro.  If it is
>>> Rocks,
>>> you shouldn't have any difficulties adding different HW.
>>>
>>> Adding different types of  nodes to any cluster comes down to how much
>>> does
>>> the heterogeneity of hardware affect applications. For example, Rocks will
>>> happily integrate 2.33 GHz Dual Core xeon and a 2.88GHz quad-core xeon
>>> node
>>> into the same cluster.  Does the clockrate difference matter to your
>>> applications. How about the #cores/node.  two four-core nodes !=1
>>> eight-core
>>> node.   The difference might matter for an application spanning the nodes.
>>>
>>>
>>>
>>>
>>>
>>>
>>>       
>>>>  I am just attempting to get a view of the ups-n-downs of the possibly
>>>> adding more compute nodes but from a different vendor.  From my other
>>>> experiments on my crash-n-burn test cluster - I have been able to add a
>>>> variety of different machines - all running redhat enterprise 5 with
>>>> oscar,
>>>> but I am not sure with CENTOS with rocks.
>>>> Thanks for any and all views.
>>>>
>>>>
>>>>         
>>> Rocks has handled heterogeneous hardware from day 1.
>>>
>>> -P
>>>
>>>
>>>
>>>       
>>>> --
>>>>
>>>>
>>>> Joseph Norris
>>>> Programmer III/System Administrator
>>>> UC Merced School of Natural Sciences
>>>>
>>>> Phone: 209-228-4576
>>>> Cell:  209-201-3410
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>         
>>>
>>>
>>>       
>> --
>>
>>
>> Joseph Norris
>> Programmer III/System Administrator
>> UC Merced School of Natural Sciences
>>
>> Phone: 209-228-4576
>> Cell:  209-201-3410
>>
>>
>>
>>     
>
>
>   

-- 


Joseph Norris
Programmer III/System Administrator
UC Merced School of Natural Sciences

Phone: 209-228-4576
Cell:  209-201-3410


[prev in list] [next in list] [prev in thread] [next in thread]