back to article Superdome X gets pumped on Skylake to become Superdome Flex

HPE has upgraded its Superdome X server to use Skylake CPUs, bragging that it's the world's most scalable and modular in-memory computing platform. Superdome is HP's line of mission-critical, highly available servers that used Intel Itanium processors and have moved to using Xeons. The Superdome X takes x86 Xeon E7 v4 CPUs. …

  1. Jim 59

    6TB if memory? That's more than every Sinclair spectrum manufactured, put together. In fact it might be more than every 8 bit home computer in the UK.

  2. Anonymous Coward
    Anonymous Coward

    Superdome in name only

    I was reading the article and wondering: why do those specs sound so familiar, as if I have seen them before? And then I remembered: Hewlett Packard Enterprise acquires SGI.

    It is nice to see that the SGI UV line lives on - even if the name is gone.

    1. Korev Silver badge
      Coat

      Re: Superdome in name only

      I see you're trying to Fuel some high Octane rumours about their Origin...

      1. Korev Silver badge
        Pint

        Re: Superdome in name only

        There's one of these for anyone who could accept the Challenge to squeeze a Tezro pun in -->

      2. ToddRundgrensUtopia

        Re: Superdome in name only

        That would be a Challenge Korev

        1. Korev Silver badge
          Coat

          Re: Superdome in name only

          Yep, we've got a long Tezroad ahead

          1. ToddRundgrensUtopia

            Re: Superdome in name only

            Follow the Indigo road

            1. ToddRundgrensUtopia

              Re: Superdome in name only

              To the Onyx city

      3. big_D Silver badge
        Coat

        Re: Superdome in name only

        Wouldn't that be high Optane rumours?

    2. ToddRundgrensUtopia

      Re: Superdome in name only

      I don't think it is an SGI UV. They were dual socket chassis with NumaLinks between. This is a modular Superdome, (originally Convex), which is a ccNUMA like SGI UV, but with much wider memory bandwidth and therefore much more expensive.

      1. curiiousguy

        Re: Superdome in name only

        It is a SGI UV300, I am an ex-SGI field engineer and had installed it in couple of sites before HPE acquisition.

        1. ToddRundgrensUtopia

          Re: Superdome in name only

          Having checked again, it's a Superdome blade design, so the RAS stuff from Superdome, but using the SGI interconnect instead of carrying on development of the Superdome, (dragonhawk), ccNUMA chipset.

          Certainly makes sense to concentrate on one.

  3. Steve Davies 3 Silver badge

    The cost?

    high, high and you what? you can't be serious.

    Still it will be less that trying to license Windows + Oracle + SAP on it. :) :) :) :wink:

  4. Alistair Silver badge
    Windows

    I've been

    Waiting for these. We have a target environment that needs this sorta toy - and some budget next year for it......

    1. ToddRundgrensUtopia

      Re: I've been

      Alistair,

      There is a much cheaper option for running large in memory workloads such as Hanna and it doesn't come from a company that might not be here in 5 years time.

      1. ManMountain1

        Re: I've been

        I accept your comment in that hardware vendors might be expected to have a hard time in coming years, but I can't see HPE disappearing up in smoke before any others do. It may have got smaller but the hardware bit of it is still ticking along and still generating a tidy profit.

  5. seven of five

    Which technology is used to combine the individual servers?

    HP won´t tell me, apart from a "scale activation kit" Q2N14A, which interconnect is used.

    Schould be PCIe, to enable remote memory, shouldn´t it? Is there more info on this?

    1. ToddRundgrensUtopia

      Re: Which technology is used to combine the individual servers?

      It can't be PCIe as it's not cache coherent. SGI's is NUMAlink and Convex licensed chips from Dolphin Interconnect (Norwegian firm). Think Intel's QPI but over many more nodes.

      1. seven of five

        Re: Which technology is used to combine the individual servers?

        Found it, it is the ports on the lower front side. With 7x4 ports, this will be point to point for all sockets. Could be really quick, @ElReg: can we please have some more info in this?

  6. Anonymous Coward
    Anonymous Coward

    Scale-out?

    SAP Hana and analytics are scale-out clustered workloads. Hana runs fine on a cluster, as it is scale-out. Hana is for analytics, and runs fine on lot of nodes in parallel. Then we have the standard SAP, which is not for analytics, and that SAP runs on non-cluster, scale-up server for OLTP storage of data. OLTP SAP = non clustered, scale-up. HANA = clustered scale-out workload. SGI never had large scale-up non clustered servers (the biggest was 8-sockets or so?). SGI had large scale-out clustered servers such as UV2000, Altix, etc - with 10.000 of cores.

    So, this can run Hana clustered workloads well, bu I wonder if this can run non clustered workloads well? Such as ordinary standard SAP? What are the SAP scores? The Fuijtsu SPARC M10-4S with 24-sockets (?) has around 840.000 SAPS. Or was it 32-sockets? I dont remember. But Fujiutsu M10-4s scales up to 64-sockets.

    1. Anonymous Coward
      Anonymous Coward

      Re: Scale-out?

      Theres not a single benchmark published with Superdome Flex besides the many press claims, so my guess is that its performance isn't that great. Although HPE makes it sound like a scale up system, the only real part of it being scale-up is that it runs a single OS across the clustered boxes. Its essentially a glorified clustered solution. Any real scale-up Database like Oracle Database and its in-memory option probably won't even run on this system as the system bandwidth and especially bisection bandwidth across all the nodes must be horrible especially at scale. And what about JVM performance to run BigData/Analytics type workloads? Don't see any benchmarks there either. There is talk about an SAP SD 2-Tier result but I don't see it published. The previous fastest Superdome X with 16 x Xeon E7-8890 v4@2.20 (384-cores) delivered 644,830 SAPs. Assuming that the Xeon Platinum 8180 @2.50Ghz is ~20% faster/core than Xeon E7-8890 v4 @2.2GHz (that’s what the HPE Dl580 Gen 9 vs 10 shows) and assuming the SD Flex scales similarly, a 16 x Xeon Platinum 8180 (448-cores) should achieve around 900K SAPs (that’s ~2K SAPs/core). Looking at the SAPs benchmark, the leader is currently the SPARC M7-8 (256-cores) @ 713K SAPs. Oracle recently launched the SPARC M8 which appears to be ~40% faster/core, so an SPARC M8-8 should scale to well over 920K SAPs. So an SPARC M8-8 will easily outperform a Flex 16-socket with ~1.8x faster/core performance. That’s quite a considerable difference when you license the Database on per core basis! And the Fujitsu SPARC M12-s, scaling to 32 x sockets (384 x cores) looks to be about another 25% faster/core vs SPARC M8 so this system looks to still be the leader across all systems!

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019