locked
Hyper-V and NUMA RRS feed

  • Question

  • I recently saw a presentation on Hyper-V, which I thought was going to be more technical than it was. I did not have the opportunity to ask, but I will here. How does Hyper-V handle logical SMP scheduling, especially in relation to a NUMA architecture, such as Opteron Quad-cores? Is there any optimization built-in, and if so, are there directives to control that optimization (i.e. enable/disable)?

    Thks, - clikdude
    Tuesday, April 15, 2008 4:29 AM

Answers

  • Hyper-V is NUMA aware, and will attempt to schedule threads on processors with local memory to avoid crossbar memory access.

    Thursday, April 17, 2008 2:08 AM

All replies

  • Hyper-V is NUMA aware, and will attempt to schedule threads on processors with local memory to avoid crossbar memory access.

    Thursday, April 17, 2008 2:08 AM
  • However, it must be admitted that Microsoft is in its infancy with writing NUMA aware oerating systems. In addition NUMA is limited by the Intel architefcture. About the only company that can do NUMA worth a darn when using an Intel platform is IBM. They had to create their own chipset to do so, they only use the Intel processors, and, they use an expensive, high-speed interconnect between the chasis. Once you go to that level however you are well over $100,000.

     

    So the real question is, why would you want to spend over $100,000 on an Intel based system. At that price point I start steering clients away from Microsoft and Intel platforms and more towards proprietary Unix systems (Sun or IBM). There are a multiltude of reasons for doing so. You have to remember that Microsoft is limited in a lot of things becuase of the limitations of the Intel architecure. This isnt a slap at Intel either. When Intel started they never imagined that there processors would one day be used by clients demanding things such as NUMA, falut tolerance, virtualization, etc. All of that stuff already existed in the mainframe world.

     

    I think it is safe to say that high quality NUMA using Microsoft and Intel is five years away.

    Thursday, April 17, 2008 3:05 AM
  • Once you go to that level however you are well over $100,000.

    You must be kidding? :-)

    The x3950M2 (scalable server) is (probably?) slightly more expensive than the 3850M2 (up to 4 socket). Notice also that the 3850M2 can be upgraded to a 3950M2 by simply enabling the scalability ports just plugging a little circuit onto the motherboard.
    So we build-in scalability into the 3850M2 and give it away for free (although not enabled) ... how can it be a "well over 100.000$" technology? Both the 3850M2 and 3950M2 are aligned to market prices.... there might be situations where we come out cheaper than competitors and others where we come out more expensive.... but we are talking a few percentage ..... not 3x as you might suggest in your post.

    I have no interest in selling you (or other readers) our boxes on this forum ..... but I had to set the record straight for the sake of the discussion.

    Massimo.

    king@it.ibm.com.NO_SPAM
    Monday, September 29, 2008 12:09 PM