ScaleMP enables virtual Opteron systems with 512 sockets, 64TB RAM

Very large multiprocessor systems have traditionally been breathtakingly expensive propositions, and they’ve arguably become more of a niche product as commodity 2P servers have added cores and capability over time. Some problems still require a Very Large Computer, though. A company called ScaleMP has cooked up an intriguing virtualization-based solution that allows a large SMP system to be built out of an array of inexpensive servers. Today, ScaleMP and AMD have announced a collaboration that allows Opterons to act as the building blocks for such virtual systems. From the press release:

Cupertino, Calif. – September 20, 2011 – ScaleMP, a leading provider of virtualization solutions for high-end computing, today announced a strategic collaboration with AMD for vSMP Foundation to support AMD OpteronTM processor-based servers. The collaboration enables users to scale AMD-based servers beyond four processors and create AMD Opteron processor-based systems with up to 512 processors and 64TB of memory . vSMP Foundation will support the current AMD Opteron 6100 Series processor as well as the upcoming AMD Opteron processors, code-named “Interlagos”, leveraging the new x86 “Bulldozer” core.

So yes, up to 128 quad-socket servers can participate in one virtual system. ScaleMP says no custom hardware or chipset is required, yet “the vSMP architecture provides cache coherency, shared I/O and the system interfaces (BIOS, ACPI) , which are required by the OS.” The interconnect of choice is Infiniband, as one might expect.

ScaleMP previously supported only Xeon processors, so this update brings the Opteron up to par with the competition on this front. We’ll be sure to grab a rack full of Bulldozer boxes for some SPECjbb testing ASAP.

Comments closed
    • internetsandman
    • 8 years ago

    *Obligatory ‘can it play Crysis’ post*

      • ronch
      • 8 years ago

      Oh boy.

    • bcronce
    • 8 years ago

    One would really have to make programs NUMA aware with this system, as locality of data would be crazy important.

    edit: Effectively you have a distributed system showing as a single system. NUMA data locality issues are now more important and your program’s memory allocation should still “act” as if they’re distributed.

      • ew
      • 8 years ago

      I’ve never had a chance to work on a system like this but I suspect that cache oblivious algorithms would be effective on NUMA systems too.

      [url<]http://en.wikipedia.org/wiki/Cache-oblivious_algorithm[/url<]

        • bcronce
        • 8 years ago

        Nifty read. +1 from me

    • Jigar
    • 8 years ago

    [i<]We'll be sure to grab a rack full of Bulldozer boxes for some SPECjbb testing ASAP.[/i<] So do you have it handy at damage lab ?

      • ronch
      • 8 years ago

      Why not ask DonanimHaber? They’re the Bulldozer experts!

    • evilpaul
    • 8 years ago

    Nobody cares about server Bulldozer!!! (OK, well some people do, but we require moar desktop info!)

    • dpaus
    • 8 years ago

    Queue obligatory question about Crysis in 3…. 2…… 1….

    This is an interesting development, especially for pharmaceutical companies and universities, as it brings Cray-level capabilities to commodity hardware.

      • kamikaziechameleon
      • 8 years ago

      you mean crysis?

        • dpaus
        • 8 years ago

        Ooops, just gave away what an avid gamer I secretly am….

          • BiffStroganoffsky
          • 8 years ago

          You can have the secret back by sending me $50USD.

Pin It on Pinterest

Share This