Differences between revisions 3 and 5 (spanning 2 versions)
Revision 3 as of 2008-06-27 16:44:01
Size: 3369
Editor: GlennTarbox
Comment: reorganization
Revision 5 as of 2008-11-14 13:42:00
Size: 3417
Editor: anonymous
Comment: converted to 1.6 markup
Deletions are marked like this. Additions are marked like this.
Line 1: Line 1:
## page was renamed from DsageNg/InfinibandSmp
Line 59: Line 60:

80 Core, 80 GB SMP for $50K

This has been coming for quite a while... but looks like things might really be happening.

It appears as though InfiniBand is ready for prime time. The architecture proposed here is one where you nail up underneath the memory infrastructure, and use infiniband to integrate multiple blades into one big SMP machine.

The idea is you run one SMP OS image with however many cores, and there is no change needed to the OS at all. The memory appears to the OS as one contiguous block and other fanciness is used to support cache coherence etc. I'm sure additional black magic is needed...

This architecture, apparently, has been proven and is working at various places. The key to the activity described below and in the attached docs is making it available to the unwashed masses (guys like us)

Clearly, more investigation is needed.

However, the person / company in question has the hardware and is testing... they're going to have a testbed up in a week or two and would be willing to give us shell access to compile Sage and see how it works on a class of problems.

I think the idea is to run some Python Processing task with as many processes as CPU's and see what happens.

So, rough numbers, for:

  • rack of 10 blades
  • each blade with 2 CPU/s = 8 cores
  • with 1 GB / core = 80 GB
  • would cost somewhere just over $50K

Thats 80 Cores and 80 GB for $50K fully SMP. Hmmm...

One more thing... wanna double it? Just buy another rack, included in the price above is an InfiniBand Switch with, I believe, 24 ports... nail it together and you have 160 cores and 160 GB for 2X the cost... basically doubling processors in SMP configuration for slightly less than double the cost.

The email below says more. The documents and links refer to IBM stuff... but this works in general and I believe that the platform we'd test on would be SuperMicro...

Hi Glenn,

Great talking with you today.  As promised, here is a high-level white paper on the vSMP arch, and some more customer-facing documentation.  In the blade space, it looks like it presently works with a two-socket, Intel Seaburg chipset blade, for a total of 8 cores and up to 64GB per blade.  The Seaburg chipset gives access to the 1600MHz procs and 800MHz memory, so that fits well with apps. that are bound by cycles and FSB memory transfers. 

Currently, this thing can be packaged in the following ways:

-Blade: (virtual 20 socket system with 640GB memory in 7U)
--Up to 10, dual-socket blades with up to X5482 procs (3.2GHz, 12MB, 1600FSB)
--Up to 64GB per blade.  
--IB HCA mezzanine cards, and IB switch in chassis.

-Twin: (virtual 4 socket system with up to 128GB* in 1U) *uses 8GB sticks

-Rack Server: (virtual 4 socket system with up to 128GB* in 2U) *uses 4GB sticks

Refer to scalemp.com for more info, and specifically to: http://www.scalemp.com/vfs-specifications/ for their hardware qual. list.

There are 2 attachment(s) stored for this page.

  • [get | view] (2008-06-25 22:40:33, 764.7 KB) [[attachment:TheVersatileSMP(vSMP)Architecture-v1.5-WhitePaper.pdf]]
  • [get | view] (2008-06-25 22:41:21, 3024.5 KB) [[attachment:vSMPFoundationEndUserv1.2-2008-05-23.pdf]]
 All files | Selected Files: delete move to page copy to page


DsageNg/HardwareArchitecture/InfinibandSmp (last edited 2008-11-14 13:42:00 by anonymous)