November 02, 2011
Now that the latest AMD "Interlagos" Opterons and Intel "Sandy Bridge" EP Xeons have begun shipping, at least for volume deployments, Appro has announced support for the latest x86 CPUs in its upgraded Xtreme-X HPC line-up. The new systems will soon be appearing in supercomputing centers in the US and elsewhere.
Besides the addition of the "Interlagos" CPU (aka Opteron 6200) and "Sandy Bridge" EP processors (aka Xeon E5), Appro is also including support for Intel's Many Integrated Core (MIC) coprocessor and FDR InfiniBand. The new Xtreme platform will also retain support for the current NVIDIA GPU modules, namely the M2090 Teslas.
The new Xtreme platform uses the GreenBlade building block, which comes in a 5U subrack configuration for a 10 node set-up and a 8U subrack for an expanded 16-node version. The 5U subrack is employed for the two-socket motherboards, while the 8U enclosure houses the four-socket version. According to Maria McLaughlin, Appro's Marketing Director, the 8U subrack is a bit of a special case, since most of their customers opt for the traditional two-socket nodes. If accelerators are desired, they are matched to their CPU counterparts in a 1:1 ratio -- 5 CPU nodes with 5 GPU/MIC nodes or 8 CPU nodes with 8 GPU/MIC nodes.
The only mainstream accelerator left unloved is AMD's FireStream GPU, mainly due to a lack of enthusiasm for the products by HPC customers. Other than that, Appro has maintained its vendor-neutral stance as far as microprocessors go, allowing customer to mix and match CPUs and accelerators at will. So, for example, AMD Opteron CPUs could co-mingle with Intel's MIC coprocessor in the same system. Theoretically, a system with CPUs, GPUs, and MICs could be constructed, as Dell is doing for TACC with their Stampede supercomputer.
Of course, the MIC won't be an official product until "Knights Corner" is launched in late 2012 or early 2013, but according to McLaughlin, they already have a customer who is deploying an Xtreme-X cluster equipped with pre-production (Knights Ferry) Intel MIC coprocessors. Although Appro didn't say so, the system in question is likely the one at Sandia National Labs, which has an Appro-based testbed system there with pre-production MIC chips.
The Sandia system isn't the only pre-launch Xtreme-X in the field. Appro has been collecting orders on the new platform for months and has already shipped a number of machines. These include the initial HPC clusters for the new Tri-Lab Linux Capacity Cluster 2 contract announced back in June. All of these are Xeon E5-based systems.
Another E5-equipped machine is the 800-teraflop supercomputer installed at the University of Tsukuba's Center for Computational Sciences. Most of the flops here are courtesy of NVIDIA GPUs though, with each node pairing up two Xeon CPUs with four Tesla modules. The supercomputer will support particle physics, astrophysics, biophysics and climate research, and represents one of Appro's first capability systems they've installed.
The Gordon system at the San Diego Supercomputer Center (SDSC) is yet another Appro Xtreme-X machine using the new Xeon E5 processors. Gordon is a 200-teraflop cluster aimed at data-intensive science workloads, which, is enabled primarily by the use of large memory virtual SMP nodes and flash memory. Regarding the latter, Gordon is being outfitted with 300 TB of Intel's iSolid-State Drive 710 devices.
They won't all be Intel systems though. McLaughlin says a large Interlagos-based Xtreme-X machine is about to be deployed for a US government agency, although the announcement is not yet public. Beyond that, Appro is still getting some mileage out of the previous generation 6100 Opterons. In September, Los Alamos National Lab installed Mustang, a 353-teraflop Xtreme-X cluster with the 12-core "Magny-Cours" processors.
Even though FDR InfiniBand is now supported on the platform, all of the systems described about are using QDR, although McLaughlin says some deals with FDR systems are in the works. In any case, Appro will continue to support both InfiniBand speeds, as well as Gigabit Ethernet for systems where lots of bandwidth and low latency is not a big deal.
The new Xtreme-X platform has encouraged Appro to start thinking more about capability systems, like the Tsukuba super mentioned above. Up until recently, the company has been singularly focused on capacity clusters, that is, systems that are partitioned to run a number different workloads simultaneously over relatively short periods of time.
Capability systems, on the other hand, typically run one large application, or at least one large application at time. And the codes -- things like climate modeling, physics simulations, seismic processing, and certain types of financial risk modeling -- can run for days or even weeks.
Such systems require more fault tolerance, redundancy and failover management, not to mention more computational power. Although this level of HPC is a bit of stretch for the company, capability machines usually entail higher margins than capacity systems, so Appro has been busy beefing up its software stack and hardware to support the needed features. With the newer faster processors, FDR InfiniBand, and the more mature software stack, the company believes this is a good time to move up the food chain.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?