[Beowulf] Nehalem memory configs

[Jason Clinton]

On Wed, Apr 8, 2009 at 2:49 PM, Peter Kjellstrom <cap at nsc.liu.se> wrote:

> Can anyone confirm that mixing different sizes of dimms, even when keeping
> to
> a by-three-symmetric configuration, actually does degrade performance?
>
> That is, first hand information that config 1 will be slower than config 2:
>
> 1: two sockets, 3x 1G and 3x 2G per socket RDIMM (total 18G @1066)
> 2: two sockets, 6x 1G(or 2G) per socket RDIMM (total 12G (or 24G) @1066)


No, it does not degrade performance as long as the CAS latencies of both
DIMM types are identical and you would not get a clock-down by using that
quantity of RAM ranks. We have been able to attain Intel's reported, magical
35GB/s number using ICC. GCC 4.3 is around 31GB/s. (At 1333MHz.)

We're a little reticent to approach this, though. It's certainly easier to
explain whole units of divisible RAM quantities to customers. The whole
tri-channel memory thing is already hard enough to have a conversation
about. Especially when people write their specs as X GB of memory per
core--where X is power of two.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.scyld.com/pipermail/beowulf/attachments/20090412/d0a79088/attachment.html