Press Conference

CNRS CHOOSES THE IBM POWER4 SUPERCOMPUTER
FOR IDRIS, ITS NATIONAL COMPUTING CENTER

Paris - February 6, 2001


UPGRADING THE IDRIS COMPUTING RESOURCES

Since it was set up in 1993, IDRIS has been managing an intensive computing environment that is highly advanced, versatile, and highly upgradable. This environment responds to the need to contribute as effectively as possible to the excellence of scientific research requiring sophisticated computing resources. For IDRIS, scientific computing is a tool that serves numerical simulation, and most of our efforts are directed towards technologies that are liable to have a major impact on national research as a whole.


Current technological issues

Current scientific computing, and indeed all computing, has been profoundly marked by the arrival in force of microprocessors, which began in the 1980s and was consolidated in the early nineties. This, coupled with the end of the Cold War, which brought with it a major reduction in defense spending in the USA, marked a downturn for technologies dedicated to very high performance scientific computing.

The very strong trend towards microprocessors using CMOS technology has facilitated the development of general machines based on mass-production technologies that are more cost effective since they have markets that are wider than the scientific market. The term “COST” (“Commodity Off-the-Shelf Technology”) is used to designate the currently dominant technologies. All of the hardware used today for scalar supercomputers has been designed for generic use in a wide variety of systems.

This approach, together with the development of multimedia and of the Internet has led to computer science touching upon every aspect of daily living. However, the transformation is less remarkable for very high performance scientific computing. Admittedly, we have been witnessing a spectacular increase in the computation power of microprocessors for over 15 years now. The constant reduction in the size of transistors has been offering ever-improving large-scale integration possibilities, and enabling both the number of transistors per chip and their working frequencies to be increased. This has led to an exponential increase in their computation power, which has been doubling every 18 months. However, the term “spectacular” is appropriate only if the data is already in the guts of the microprocessor. When the microprocessor has to retrieve the data from the memory – which it almost always does – it is quite a different story. Over the last 15 years, memory access performance has changed very little, and the speed of data transfer between memories and processors available using COSTs has become totally unsuitable.

The only exception to this rule comes from vector supercomputers. As their name suggests, they are specially suited to processing vectors, i.e. long regular tables. Vector computing is actually merely a particular example of parallel computing. Vector machines draw their performance from two elements: processors that are specialized in vector processing, and exceptional memory-processor data transfer speeds, enabling the processors to be fed as quickly as they produce their results. Since a number of scientific codes use regular data structures, these computers have traditionally had a major impact on scientific computing. Between 50 and 60 percent of the scientific projects in progress at IDRIS use this type of architecture.


IDRISÕs strategic approach

Replacing our pool of machines to meet the scientific challenges of the new decade took place in 1999 and 2000 against the technological background we have described above. In tune with its objectives, IDRISÕs strategy has been to seek exceptional and sustained performance through architectures that are well-balanced, robust and tolerant of breakdowns, and that are equipped with mature and reliable system software capable of coping correctly with the conflicts that arise from a research instrument being shared by hundreds or even thousands of user scientists.

Needs on such a scale require exceptional machines. The need for high-quality master software to drive the computer and arbitrate the conflicts, and the need for good performance in inputting and outputting huge masses of data are difficult to satisfy with COST architectures such as clusters of PCs, which are relatively precarious. Admittedly, such architectures are very affordable, but what counts is not only the cost of an architecture but also the quality and the quantity of the scientific results obtained. From this point of view, the IDRIS machines are probably 100 times more expensive than a good cluster of PCs, but they are also over 100 times more productive: they process over 400 scientific projects that it would be impossible to process on 100 PC clustersÉ

The other major element in the IDRIS approach results from the need to predict the future in a field undergoing fast technological transformation. Thus, our main objective is not to find an immediate solution or response to the needs of the moment, but rather to identify technology capable of carrying forward the development of high-performance intensive computing.

At the end of 1998, and after in-depth forward-thinking by the Scientific Council of IDRIS, the replacement of the pool of supercomputers was launched. The objective was to increase tenfold the computation power installed on the site, both for scalar computing and for vector computing. It was deemed to be absolutely unavoidable to reinforce vector computing. It then appeared to be essential to install a new-generation scalar architecture to replace the Cray T3E which was becoming obsolete, in order to handle the 40 to 50 percent of applications that were ill-suited to vector computing.


Upgrading vector computing

The two vector supercomputers installed as of the founding of IDRIS, namely the Cray C98 and Cray C94, gave sterling service to research in France until they were decommissioned in February 2000 after several years of intensive operation with exceptional throughput. They were joined by a Fujitsu VPP300 vector supercomputer as of 1997, which doubled the computation power and remained in service until February 2000.

The first operation in replacing the equipment was to replace the preexisting pool of vector machines with a cluster of three NEC SX5 vector supercomputers. Inaugurated in June 2000, and currently comprising 40 processors offering 8 gigaflops each and 236 Gigabytes of overall memory, this architecture is, at present, the largest vector supercomputer in France, and the largest NEC system in the world.


Upgrading vector computing

In the mid 1990s, IDRIS moved boldly towards parallel scalar computing. At the time, it seemed that these architectures would ultimately replace vector architectures (we have now seen that this has not yet happened). A Cray T3D supercomputer was installed in 1995 (128 Alpha EV4 processors, and 16 gigabytes of memory in total), and was replaced in 1996 with the current Cray T3E (256 Alpha EV5 processors offering 600 megaflops per processor, and a total of 32 gigabytes of memory). These architectures constitute a superb technological achievement, offering a very large number of features dedicated to scientific computing. Admittedly, the T3E is based on standard microprocessors, but it incorporates into its architecture a large number of hardware elements specially designed for high-performance computing. It is situated at the peak of a technological approach which has since declined as a result of the COST approach.

After several years of very intensive operation, this platform will be decommissioned at the end of 2001 after a new-generation scalar supercomputer has been brought into service. We feared it would be difficult to replace: the performance and quality criteria adopted by IDRIS are incompatible with the sometimes precarious software environments that accompany COST clusters. Fortunately, an attractive option lay before us: at the dawn of the new millenium, the carrier of very high performance parallel scalar computing at IDRIS will be the new IBM Power4 architecture. This is because, at the beginning of this new century, IBM's strategy approach is perfectly in tune with the strategy vision of IDRIS. The Power4 project implies abandoning the COST approach: we stop designing and manufacturing generic components and microprocessors, and we work towards a new overall structure, in which all of the elements (processors, communications with the memory, communications between processors, inputs-outputs, etc.) are reconsidered and harmonized. This approach, which is very close to what we experienced at the time of the T3E, appealed to us. IDRIS has found a very worthy successor for the T3E.

This innovative architecture is capable of opening up new prospects for very high performance computing. We expect that it will bring scalar computing closer to vector computing, by offering sustained power higher than the 10-15 percent of peak performance that is characteristic of todayÕs scalar machines. The configuration that will be installed at IDRIS at the end of 2001 will be made up of 8 Power4 nodes interconnected via a Colony network. It will lead to a nominal power of 1.2 teraflops, and an overall memory of 832 gigabytes (the nominal power of the T3E is 150 gigaflops, and its overall memory is 32 gigabytes!). Each of the eight Power4 nodes is more powerful than the T3E.


21st century computing resources

The word ÒconstellationÓ was used for the first time last year at the international scientific computing conference of Dallas (SC2000) to designate a cluster made up of a limited number of hyper-powerful nodes having a large shared memory.

IDRIS is starting the 21st century with scalar constellations (IBM) and vector constellations (NEC), whose nodes have power ranging from 128 to 160 gigaflops, and a memory size ranging from 64 to 256 gigabytes. What is more, these constellations will not be totally stand-alone: developments in the field of distributed computing will make it possible to use them in harmonious and complementary ways in the near future.

These constellations are integrated in a complex environment including firstly a large number of support machines and viewing servers for high-performance preprocessing and postprocessing, and secondly a cutting-edge data-management system capable of storing up to 800 terabytes of data produced by numerical simulations.

Such an environment, accompanied by very high added-value all-level support and training services provided by IDRIS, constitutes a major asset for numerical simulation. Used in an increasingly forward-looking and innovative way, it will be capable of stimulating researchers' creativity and contributing effectively to pushing back the frontiers of our knowledge.