[Beowulf] Single threaded memory bandwidth compared between Nehalem-EP and Westmere-EX
samuel at unimelb.edu.au
Thu Oct 17 20:34:59 PDT 2013
-----BEGIN PGP SIGNED MESSAGE-----
We've got a group who bought themselves a 16 core 1.5TB Westmere-EX
system and are getting what they consider to be poor memory bandwidth
compared to an older 8 core Nehalem-EP system.
Same CPU clock speed (2.6GHz), same memory speed (1067MHz), but about
half the memory bandwidth (single threaded). I've run my own tests
with the Stream benchmark and can confirm the same.
For instance, the Nehalem-EP system typically gets around 12GB/s for
triad and the Westmere-EX system gets around 4.5 GB/s.
Now I know that I can use 16 threads (one per core) on Westmere-EX and
get far better performance, but that doesn't interest them.
I realise that the memory controller architecture is very different
between EP and EX models, but is there any inherent reason why that
should result in such poorer single threaded performance?
All the best,
Christopher Samuel Senior Systems Administrator
VLSCI - Victorian Life Sciences Computation Initiative
Email: samuel at unimelb.edu.au Phone: +61 (0)3 903 55545
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.12 (GNU/Linux)
Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/
-----END PGP SIGNATURE-----
More information about the Beowulf