European Vendors Offer Home-Grown Petascale Supers
The sinking economy has not been kind to US-based HPC vendors. SiCortex and Woven Systems have gone belly up, while SGI ended up merging with Rackable. Meanwhile, Sun Microsystems is slated to be acquired by Oracle later this year. All of these companies introduced legitimate innovation into HPC, but failed to make a go of it on their own.
As American HPC companies retrench, a new crop of European-based vendors is emerging. In our recent podcast from the International Supercomputing Conference in Hamburg, I mentioned three companies across the pond that have designed some rather interesting high-end HPC machines. Bull (France), T-Platforms (Russia), and Eurotech (Italy) all recently introduced new purpose-built HPC server platforms, although each vendor has taken a slightly different approach.
Bull, which made a relatively-recent entry into the HPC business, unveiled its new line of “Extreme Computing” (bullx) HPC systems just prior to ISC. As we reported in our original coverage, Bull added plenty of performance engineering to make sure these systems could scale out to petascale-sized systems. The system building block is a dual-socket Intel Nehalem EP blade with dual on-board QDR InfiniBand and an option for up to two GPUs for extra acceleration. A 7U chassis consists of 18 blades, with six chassis to a rack. A petaflop-sized system (sans GPUs) would consist of a mere 100 racks — about 10,000 blades.
At ISC, Bull also dropped hints of its upcoming bullx SMP platform due out next year. The company says this architecture will be able to achieve a peak petaflop with just 800 servers. Like the blades, the SMP server will also have an option for GPU acceleration. Bull is mum on the particulars of the SMP design, but a good guess is that it is going to use the upcoming Intel Nehalem EX chips. Nehalem EX can be scaled to eight sockets and up to 128 memory modules per server, which would deliver 64 cores (assuming the 8-core chip) per node and a ton of shared memory.
Also at ISC, T-Platforms was showcasing T-Blade 2, the company’s second generation blade offering. Based in Moscow, T-Platforms is a 7-year-old company, which employs around 140 people. According to the Supercomputers.ru Web project, T-Platforms has about a third of the Russian HPC market, edging out both HP and IBM. The company’s largest installed system is at Moscow State University, a 60 teraflop system based on T-Platform’s first-generation Intel Harpertown blades. A new 350 teraflop system will be installed at the university in October using the T-Blade 2 hardware. That system is scheduled to be upgraded to a 500 teraflop machine in early 2010.
Like the Bull blade, T-Blade 2 is also based on a dual-socket quad-core Nehalem, but in an even denser configuration. T-Platforms puts 32 dual-socket nodes in a 7U chassis, delivering 3 peak petaflops. And believe it or not, it’s all air-cooled. A heat sink spans each board from end to end to keep the whole thing from melting. The design uses 10 custom components, including the motherboard, memory modules, an InfiniBand switch board, and a management module, among others.
The management module is the secret sauce for the platform. It’s designed to elevate the architecture from that of typical commodity cluster to more of an MPP-like experience. The module supports a global barrier network that enables fast synchronization of jobs running on separate nodes, and a global interrupt network that reduces the influence of OS jitter by synchronizing process scheduling over the entire system. The company claims this capability allows systems to scale up to as many as 25,000 nodes.
T-Platforms also sells a line of Cell BE-based offerings (server, workstation, and two-node mini-cluster) using the latest PowerXCell 8i chip, along with a home-grown Cell compiler. And if you’re not into hardware, the company also offers an HPC on-demand service.
A somewhat similar offering to the Russian T-Blade 2 was also unveiled at ISC by Eurotech, a company based in Amaro, Italy. If you haven’t heard of Eurotech (and I hadn’t), its stated mission is to “integrate state-of-the-art computing and communication technologies into miniaturized and user-friendly solutions to improve everyday life, making it simpler, safer and more comfortable.” Up until now, the company has mostly been focused on embedded and wearable computing, but has dabbled in HPC from time to time. Check out the Eurotech Wikipedia entry for its unusual history.
As for the Eurotech super, which is named Aurora, we again find a custom-built, high-end cluster with a lot of cutting-edge technology. Like the Bull and T-Platforms offerings, Eurotech is using dual-socket Nehalem blades that can be aggregated into petascale-sized machines. The Aurora design is on par with T-Platforms for computational density, offering 3 peak teraflops per chassis.
But unlike the Russian super, the Aurora machine is water-cooled (must be the warmer Italian climate). Each blade comes with up to 160 GB of solid state disk storage for application I/O and checkpointing. There’s also something called a “programmable high performance accelerator” integrated onto the motherboard, but there’s no hint of what it actually is.
Again it looks like a lot of the innovation went into the network interconnect, which consists of a 60 Gbps 3D Torus integrated with a QDR InfiniBand network, and three synchronization networks. The 3D Torus comes with a programmable network processor if you desire more customized management of the system interconnect. If I were a network engineer, I could probably tell you what this all means, but I’ll have to leave it as an exercise for the reader.
Unfortunately, none of these interesting machines are going to be shipping into the North American market anytime soon. Bull and Eurotech will be focusing mostly on the Western European HPC market, while T-Platforms intends to concentrate on Russia and the Commonwealth of Independent States (CIS), i.e., the former republics of the Soviet Union. The Russians are also interested in finding a European partner to give them access to that lucrative market.
It’s gratifying to see HPC tech innovation occurring outside the US, especially in these economically-challenging times. It will be worth watching to see how these companies fare in their more regional markets, and if they are able to compete against global OEMs like IBM, HP, Dell, SGI and Cray. Stay tuned.