[MITgcm-support] Inter or AMD: that is the question...

Stefano Querin squerin at ogs.trieste.it
Thu Mar 21 11:23:38 EDT 2013


Dear all,

I know that there has been lots of discussions about this topic, but I  
would like to know if somebody has some new experience and/or  
information...

In a few words, we are planning new simulations using domains with  
~500x500x100 gridpoints. We are about to buy a new cluster  
(approximately 100 cores, 2GB memory per core -at least-, Infiniband  
switch, 20TB disk space) and we are facing the problem of which kind  
of processor we should choose.
I read lots of old posts and, based also on my previous experience, it  
is clear that:

- MITgcm has "heavy requirement for both memory bandwidth and network  
performance";
- too many (12?) cores per chip could lead to poor performances;
- AMD Opterons scaled better than pre-Nehalem Xeons (5-6 years ago).  
Now things seem to be more even...

I'm attaching also a -quite recent (2010)- interesting paper about  
performance issues on multicore systems. It is rather "technical" but,  
as far as I understood, Figure 4 shows that Intel-Nehalem-based  
systems are quite efficient: more than two times faster than AMD-based  
systems. So, the choice should be almost a matter of cost.
We contacted some vendors: for almost the same price we can get (for  
example) 192 AMD cores or 128 Intel cores (256 threads, if we consider  
hyperthreading...), being the other elements of the cluster almost the  
same.
In brief: same cost, about half (0.66) number of Intel cores, with  
almost double performance (or vice-versa for AMD...). So: same cost,  
same cluster performance (with Intel, probably, a bit better...).
Of course, there are many variables acting on overall cluster  
performance (hardware, software (compiler), configuration of the  
simulations...) and specific benchmarks using our present and planned  
MITgcm configurations would settle things once and for all.  
(Unfortunately not all vendors are willing to do these tests on their  
facilities...)
Anyway, generally speaking, it seems to me that, for the same budget,  
nowadays, either solution (AMD or Intel) is fine.
Do you confirm this "feeling"?

Best regards,

Stefano


P.S.: is the good performance of Intel-Nehalem cores related to  
hyperthreading?


-------------- next part --------------
A non-text attachment was scrubbed...
Name: Multicore_Systems.pdf
Type: application/pdf
Size: 540725 bytes
Desc: not available
URL: <http://mitgcm.org/pipermail/mitgcm-support/attachments/20130321/d4e2cb92/attachment-0001.pdf>
-------------- next part --------------



More information about the MITgcm-support mailing list