Less than a year after the launch of the first bullx™ models, Bull’s family of supercomputers has been extended, with a new model. But why has it been necessary to develop these different types of systems?
If you are looking for the very highest levels of performance, in every area, you need hardware that’s perfectly adapted for the job. High-Performance Computing (HPC) is no exception. Different systems meet quite different requirements, because they are optimized for certain types of use or algorithms. Which is why Bull has gradually built up a range of solutions which is now one of the most comprehensive and powerful available on the market. bullx blades – which we launched in 2009 – are exceptionally dense ‘thin nodes’, which means they are used to combine a large number of small servers within a ‘scale-out’ architecture. Then we offer a highly innovative type of blade that natively integrates GPU accelerators. This hybrid system derives directly from our earlier work, most notably Titane, the system we implemented for the French national High-Performance Computing organization GENCI, which is still the biggest GPU-based system in Europe. Most recently, we have just launched our S Series ‘supernodes’, designed for ‘scale-up’ architectures, using ‘fat nodes’. Each approach has is own specific characteristics and its advantages, and that’s why the bullx family is capable of covering the whole technological spectrum. Of course, all the different bullx models are compatible with each other, and can be brought together into a made-to-measure cluster to combine the advantages of the various architectures and meet the needs of as many applications as possible, as closely as possible.
Can you tell us a bit more about the bullx supernodes?
They are high-end nodes, quad-processor SMP servers that can be extended to 8, 12 or even 16 processors thanks to a new technology designed exclusively by Bull. In the most extensive configuration, they can form an enormously powerful processing node, with up to 128 cores and 2TB of shared memory. The bullx supernodes – which have been designed and developed by Bull’s teams working in close cooperation with the CEA – form the basis for Tera 100, the first Petaflops-scale supercomputer ever designed and built in Europe. Destined for the French nuclear weapons simulation program, Tera 100 consists of 4,300 bullx S Series servers. It features 140,000 Intel® Xeon® 7500 series cores, which makes it the largest system ever built so far using Intel® Xeon® processors. I should add that the software environment used for systems administration on Tera 100 was also developed by Bull. It is this bullx cluster suite that manages the whole system, from installation to day-to-day running. Tera 100 was powered up for the first time on 26 May 2010.
Bullx supernodes and fat nodes in general are not always reserved for Petaflops-scale systems. What are the advantages of this type of system?
In recent years, the emergence of clusters based on standard servers has really enabled the ‘democratization’ of HPC. But not all applications are designed to take full advantage of these scale-out architectures. What’s more, above a certain size, the sheer number of nodes needed to deliver the computing power that is needed can become an issue. Using fat nodes, like the bullx supernodes, can drastically cut the number of processing nodes (by a factor of 3 to 10), really simplifying the infrastructure. This not only makes it easier to implement the cluster initially, but also and more importantly the day-to-day administration of the system. Using fewer components means many fewer potential problems. And bullx supernodes provide a huge amount of shared memory. This is extremely useful for example in mesh-creation operations, in climatology, aeronautics or automotive design, crash simulations or oilfield modeling… That also goes for the mechanics of complex structures or the chemistry or physics of materials, helping to understand the behavior of components at a molecular level. As a result, many of our customers are going to be incorporating a number of supernodes into their existing clusters so, alongside their ‘thin’ processing nodes, they have access to a server with massive memory capacity.
Now that the bullx family is complete, what are Bull’s ambitions for the Extreme Computing market?
We are absolutely delighted at how well bullx has been received. Just a few months after it was launched, the readers of HPCwire voted it the best supercomputer in the world. And its commercial success is undeniable, which proves that our offering is in tune with what the market is looking for. For example, the Atomic Weapons Establishment (AWE) has just chosen bullx blades to build two 35 Teraflops supercomputers. AWE has been at the heart of Britain’s national defense for over 50 years, and we are very proud to have won their confidence. So that we can continue to meet our customers’ expectations and help them constantly push back the boundaries of innovation, we are going to dedicate our efforts to ensuring that the bullx family remains a state-of-the-art, vigorous and constantly evolving technological force. We are also going to carry on developing products and services related to Extreme Computing. We already talked about the bullx cluster suite developed by our Extreme Computing center of expertise. Bull also offers the mobull container to buy or hire; a genuine mobile Data Center which lets you deploy a new computing facility (whether traditional or HPC) in the shortest possible timescale. We also offer hosting solutions for HPC systems and sell computing power on demand. Several of our customers have handed over the management of their HPC systems to Bull, so they can concentrate on their core business or deal with peaks in business activity. Extreme Computing has been one of the foundation stones of Bull’s strategy for several years now. Bull is the only European IT maker in this sector and, with 500 experts we have Europe’s largest HPC center of expertise. Today, we are the up and coming European challenger, and we’re winning some outstanding projects. Our ambition is to be number one in Europe in this area by 2015.
The bullx S Series features two models
bullx S6010: high-end, compact processing nodes
The exclusive format of bullx S6010 nodes – L-shaped 1.5U drawers which fit together, with one upside-down on top of the other, to form a 3U drawer – enables a level of density in the 8 or 16-processor configuration (3 or 6U) never before available in this market. Specifically designed to act as compute nodes, S6010 servers feature the best available components for this function, and only those that are essential for processing.
bullx S6030: service nodes, with enhanced connectivity capabilities and storage in 3U format
bullx S6030 nodes offer advanced connectivity functions, a redundant power supply and extended storage options. This means they are especially well suited to act as management or I/O nodes that are just as effective as they are reliable.