Fujitsu received an order from the Japan Atomic Energy Agency (JAEA) for a new supercomputer system which will be the most powerful in Japan when it goes online in 2010.
The Linux cluster, serving as the core of the new system, will employ the PRIMERGY BX900 blade servers. Comprised of 2,157 nodes, the core system will have theoretical peak performance of 200 teraflops, making it the fastest supercomputer system in Japan.
The JAEA was established in October 2005 as Japan’s sole research institution in the field of atomic energy. Its research has included fast breeder reactor cycles, high-level radioactive waste processing and disposal, atomic energy safety, nuclear fusion, neutron science, and the use of quantum beams.
Until now, the JAEA has utilized two supercomputer systems: a shared system (theoretical peak performance of 13 teraflops), and a system for use in its fast breeder reactor project (2.4 teraflops). The agency’s need to process massive amounts of data has placed a heavy burden on both systems, resulting in an urgent need for the upgrade. In response, the decision has been made to introduce a new system that combines the functions of the two existing systems and can handle the massive computational demands of nuclear fusion simulations and other atomic energy research.
The new system will be used for a variety of simulations involved in nuclear plant development, including studies on fusion and fast breeder reactors, as well as assessing nuclear facilities’ earthquake resistance. It is expected to contribute to the achievement of safer atomic energy and help address the problem of global warming.
The new system, built around a large-scale parallel computation cluster, will be a hybrid system consisting of three computational server subsystems, each intended for different purposes. Reaching 214 teraflops, the total theoretical peak performance for these three subsystems is approximately 14 times greater than the present system, which will make the performance of new system the highest in Japan.
Large-scale Parallel Computation Cluster (theoretical peak performance: 200 teraflops) system will be Japan’s largest Linux cluster, employing PRIMERGY BX900 blade servers equipped with Intel Xeon X5570 processors (2.93GHz), for a total of 2,157 nodes and 4,314 CPUs (17,256 cores). Nodes will be connected using the latest InfiniBand QDR high-speed interconnect. This subsystem will primarily be utilized in large-scale computing tasks, such as nuclear fusion simulations, in which one job requires at least 100 teraflops.
Next-generation Code Development Unit (theoretical peak performance: 12 teraflops) will include a cluster of FX1 high-end technical computing servers, each equipped with SPARC64 VII quad-core processors developed by Fujitsu, for a total of 320 nodes and 320 CPUs (1,280 cores). The FX1 will primarily be used for application development for a peta-scale computer.
Shared Memory Server (theoretical peak performance: 1.92 teraflops) is configured using a SPARC Enterprise M9000 UNIX server equipped with SPARC64 VII processors. It will primarily be used in simulation computations for developing fast breeder reactors.
The three subsystems are tied together using Parallelnavi, Fujitsu’s HPC middleware, which allows them to be managed as a single system and ensures a high level of operability. Other equipment includes a 36-unit, large-scale ETERNUS DX80 disk array storage system, providing 1.2 petabytes of storage space.