IDRIS is one of the two existing supercomputing centres in France. Today, the centre hosts over 500 scientific projects and offers services to more than 1800 users in the country. IDRIS is managed by the Centre National de la Recherche Scientifique (CNRS), a French organisation for fundamental research. The second centre is called CINES and is controlled by the Ministry of Universities. Both institutions closely work together to provide the French research community with the required high performance resources.
IDRIS delivers the computational power for processing in nine different scientific domains, including environment and climat modelling; computational fluid dynamics; reactive flows; astrophysics and geophysics; electromagnetism and plasma physics; applied mathematics and theoretical physics; complex molecular systems and biology; quantum chemistry; and material sciences. The bulk of computational (27%) and data management (80%) facilities is reserved for environment and climat research. At IDRIS, 45 computer scientists and 15 technicians are working, with 18 expert engineers active in the user support team for code coupling and metacomputing.
Before November 1999, IDRIS had one scalar Cray T3E system and three vector machines at its disposal, being the Crays C98 and C94 and a Fujitsu VPP300. The systems were linked to a up to 800 terabyte supporting file server for data management via a HIPPI network. Since 60% of the applications which run at IDRIS, preferably rely on vector processing, the need emerged to substitute a cluster of 3 NEC SX5's to the old vector machine. As such, a 16-processor NEC SX5 with 128 GB of memory was implemented in November 1999. Two others followed in April 2000, another 16-processor with 64 GB of memory and a 6-processor with 32 GB of memory. The last one will be extented to 8 processors.
At the same time, a new IXS interconnect network has been installed with a total disk space of 5 TB. The T3E scalar system will also be upgraded whereas a novel scalar architecture with teraflop peak performance will be available by the end of 2001. As for the NEC SX5 constellation, each processor has a peak performance of 8 Gigaflops. However, 25% of the 150 migrated vector codes currently in action only achieve a sustained performance of 1 Gigaflop. This is due to the fact that both the molecular and hydrodynamics codes, and the particle methods for plasma physics rely on indirect addressing and scatter-gather operations which are a rather weak point for the latest CMOS Japanese vector systems.
According to Professor Alessandrini, the rest of the vector codes shows a higher sustained performance of 2 to 6 Gigaflops per processor with a peak of 3 Gigaflops per processor. Gang scheduling has been intensively tested on the NEC platforms because of the necessity to mix single processor and parallel jobs on the very same node. The results were quite satisfactory. In addition, shared memory processing with OpenMP and Posix Threads has been distributed among the scientific users to gain more experience with these types of models.
The IDRIS staff has planned to use both big nodes for batch processing, whereas the smallest node will serve for interactive programming and experiments in metacomputing and code coupling. In this regard, these jobs will not act as an extra burden on the system's principal workload. The scientific users however will have the pleasant experience of dealing with a 40-processor machine. Professor Alessandrini has the impression that the future looks bright for IDRIS, given the promising mixed programming model paradigm, which will enable the supercomputing centre to explore innovative ways of operating.