Mark Hahn wrote:
there's a company, ScaleMP, which seems to be selling some kind of kit
which enables to fairly large shared-memory x86_64 systems.
their website is nearly useless (http://www.scalemp.com/), but a little
more info can be had from SGI, which apparently uses ScaleMP for their
f1200 product (rebadged Ciara?).
they claim to support up to 32 sockets, 512GB memory. "Versatile SMP".
The signage at SC07 claimed up to 1 TB ram. Works with Intel and AMD.
SGI seems to aim it purely at structural/cfd/crash sims - mainly using
Abaqus and related tools.
Yup. Big single memory/system image machines.
as far as I can tell, the 32s configuration is 4x 8-socket boxes,
each with 7x 1 Gb links to their peers. seems to claim that it runs
unmodified rh/fc/suse systems. marketing docs claim that the secret
sauce is bios firmware, and mention that "at least 10%" of the memory
Yup. This is about right.
is "reserved" for system cache. I'm not sure how the bios is involved,
but it sounds like a pretty generic network-shared-memory system,
which would be OK for uncontended pages, but would thrash once procs
wait for ~80 us to reference a remote page...
They really want you to use a fast network connection ... think IB or
similar.
This is BTW quite similar to what Panta was doing when it was alive.
I would be most grateful if anyone has experience or knowlege of what
ScaleMP actually does.
Build big SMPs out of little SMPs. Shared memory, NUMA, need to
localize pages and access. But you can allocate several hundred gigs if
you want/need to.
The boxen are commodity. Shai Fulthem (CTO) occasionally posts on LKML
and other places (may lurk here). Technology is neat, allows you to
aggregate smaller units into larger units. Not on the fly, but on boot
(cold boot).
--
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: [EMAIL PROTECTED]
web : http://www.scalableinformatics.com
http://jackrabbit.scalableinformatics.com
phone: +1 734 786 8423
fax : +1 866 888 3112
cell : +1 734 612 4615
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
