> Kilian CAVALOTTI wrote:
>> On Wednesday 08 April 2009 19:48:30 Joe Landman wrote:
>>> As an FYI, Beowulf veteran Jeff Layton wrote up a nice article on
>>> memory
>>> configuration issues for Nehalem (I had seen some discussion on this
>>> previously).
>>>
>>> Link is here:
>>> http://www.delltechcenter.com/page/04-08-2009+-+Nehalem+and+Memory+Configur
>>> ations
>>
>> That's a great write-up, gathering all the useful information in one
>> place. A must read. Thanks Jeff!
It is a good write up, very good ... but there looks to be an error under
"Memory Bandwidth
Performance" ... first, for socket-local-memory, I am not sure of the relevance
of
the QPI GT rate. I would calculate the maximum rate to socket-local-memory
with 1 DIMM per channel at:
1.333 GHz x 3 Channels x 8 bytes = 31.992 GBytes/sec
not 35 GByters/sec. Maybe I am doing something wrong. Rates to socket-remote
memory would be lower of course, bounded by the QPI 6.4 GT/sec transfer rate,
which
on this 2 byte duplex channel would be 25.6 GBytes/sec I believe. There would
also
be a performance drop due to the latency increase from 60 nanos to 100 nanos.
Perhaps I have made an error somewhere ... if so, I would love have someone
tell where I
made my mistake.
rbw
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
