On August 17, 2016 11:55:16 PM CDT, Pontus Pihlgren <pontus at Update.UU.SE> wrote:
On Wed, Aug 17, 2016 at 11:01:43PM -0500, Chris
Elmquist wrote:
The SGI shared memory systems, the most current of which is called
"UV", is a NUMALink interconnected design. They have a number of
ASICs that implement the interconnect and protocol and bridge to the
Intel processors intra-processor link, which is called QPI.
I see, and I guess that NUMAlink is a decendant of the interconnect
used
in SGI Origin2k and 3k systems?
That's right. It took them a while to transition from MIPS to x86 with this
technology and so we did the MPP cluster stuff as an interim product around 2007.
The SGI
cluster products use Infiniband for interconnect.
And to take advantage of that you would use MPI?
Yes. Which sits on top of OFED (nee OpenIB) managing the fabric.
Chris
--
Chris Elmquist