[gmx-users] dual-core performance

[Erik Lindahl]

Hi,

On May 3, 2005, at 8:49 AM, David van der Spoel wrote:

> On Mon, 2005-05-02 at 23:43 -0600, Justin MacCallum wrote:
>
>> Hi,
>>
>> the new dual core Opterons have been mentioned several times on the
>> mailing list recently. We're currently investigating our options  
>> for a
>> cluster upgrade later this year. Has anyone benchmarked GROMACS on
>> these machines yet? I believe both cores share a single memory
>> controller on these chips. How does/would that affect performance?
>>
> Actually Erik Lindahl mentioned the opposite, that each processor does
> have its own memory controller. Anyway, as soon as I get the  
> machines I
> will do benchmarks. We already have normal Quad Opteron machines for
> which there are benchmarks on the website. It won't differ too much,
> since GROMACS is not very memory sensitive anyway.

Sorry, I probably confused David - here's the situation as I  
understand it:

There are significant differences between AMD and Intel dual-core  
chips. Each Athlon _die_ (i.e. physical assembly) has it's own memory  
controller, which is shared by the two cores.

However, the Opterons have the entire north bridge including a system  
request interface and crossbar switch built-in on the die, which  
means inter-core communication is potentially very fast.

In contrast, Intel dual core chips use an external memory controller,  
which will be shared by all four cores in the case of dual-dual-core  
systems. To make things worse, any communication between the two  
cores on a common die has to go over the external, shared, bus.

My hunch is that the dual-core Opterons should do rather well, but  
not the Intel ones.

It's not realistic to expect 100% improvement, though, but as soon as  
I get access to a system I'll see what we can do to tune it.

Cheers,

Erik