[Beowulf] Building new cluster - estimate

[Robert G. Brown]

On Mon, 28 Jul 2008, Ivan Oleynik wrote:

> Bill,
>
> Thank you for your comments.
>
>
>>> Yes, both opteron 2356 and Xeon E5440 are comparable in pricing (~ $700),
>>> but it is 0.5 GHz difference!
>>>
>>
>> Er, so, aren't you more concerned with performance than clockspeeds?  I've
>> seen little if any correlation.
>>
> Yes, I care about performance, but our previous experience with running our
> mpi codes on TACC computers (Ranger, Barcelona 2.0 GHz) and Lonestar (Xeon
> 5100 2.66GHz) is not in favor of AMD. They have recently upgraded Ranger to
> 2.3 GHz, I am going to run tests and report the results.
>


In both processor families there is a strong (nearly linear) correlation
with clock for obvious reasons, but they don't have the same base or
slope.  In my experience, AMDs outperform Intel at equivalent clock, by
ROUGHLY a scale factor that ranges from 1.5 to maybe 1.1 or 1.2
(depending on just what family is being compared to what family).  But
this experience is far from exhaustive, and because it is so variable it
is very difficult, as you say, to make definitive statements based on
clock alone.  In fact, the only sensible thing to do is indeed run tests
and compare on a family-by-family, program-by-program basis since
outside of simple stream-like stuff there are differences per
application and program pattern -- integer and float don't even
necessarily scale the same way.

   rgb

>
>
>>
>> Keep in mind that Intel hasn't changed their memory system much in quite a
>> few
>> generations.
>>
>
> But FSB was bumped up substantially (from 400 MHz in 2002  to 1600 MHz now).
>
>
>
>>  I am going to run some tests, but our previous experience with 2.0GHz
>>> Barcelona was not very encouraging.
>>>
>>
>> What tests did you run?  8 threads on each?  Basically if you are cache
>> friendly the intel systems win, if not the opteron systems win.  Over twice
>> the memory bandwidth can be quite handy.  AMD still wins most 8 thread
>> benchmarks (spec CPU and spec web among others) and many application tests
>> that I've done.
>>
>
> These are the MPI jobs, 8 processes per 8 cores. I am not sure about the
> setup, but I thought it is worth to disable multithreading. Correct me if I
> am wrong.
>
>
> Ivan
>

-- 
Robert G. Brown                            Phone(cell): 1-919-280-8443
Duke University Physics Dept, Box 90305
Durham, N.C. 27708-0305
Web: http://www.phy.duke.edu/~rgb
Book of Lilith Website: http://www.phy.duke.edu/~rgb/Lilith/Lilith.php
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977