[Wrf-users] Recent smpar-only benchmarks?

[Andrew Robbie (Gmail)]

On 09/03/2011, at 5:17 AM, Bart Brashers wrote:

> Has anyone done some recent smpar-only WRF benchmarking of these two
> systems:
>
> Intel Xeon X5660 (6 core, 2.8 GHz)
> AMD Opteron 6174 (12 core, 2.2. GHz)
>
> You can get twice as many cores from the AMD system compared to the
> Intel system, for about the same amount of money.

I'd like to see some numbers on this too. It actually quite possible  
to get a quad-socket
board populated with AMD 8 core (eg 6134) for less than a dual socket  
X5660 system.
So 32 AMD opteron cores for the same price as 12 nehalem ones, because  
the AMD
CPU has no price penalty for running quad socket.

However I think the more relevant limitation is the memory bandwidth  
available. Each AMD
socket addresses one four-way interleave RAM bank, and each Intel has  
a three-way
interleave bank. However, especially in the case of the AMD 12 core,  
that bandwidth is being
split among many cores. The AMD chip also has an unusual cache design  
and each
pair of cores has shared access to a pair of FPUs -- hard to guess how  
this effects real
world performance.
c.f:  http://www.realworldtech.com/page.cfm?ArticleID=RWT082610181333&p=3

> Note I'm not interested in dmpar scaling, so only up to 24 (Intel)  
> or 48
> (AMD) cores per run.  I typically have many 5-day chunks to run, so I
> use smpar only.

I think dmpar benchmarks might also be greatly influenced by other  
tunables to do with MPI.

Regards,
Andrew