Looking back, NAL started in 1981 with an M380, about 20 MFlop/s, then the VP400 1 GFlop/s, the NWT 280 GFlop/s. NAL headed the Top500 for three years with the Numerical Wind Tunnel (NWT) based on Fujitsu vector processors, starting in November 1993 and is called in the talk NSII. In the meantime NAL started multidisciplinary analysis, multiscale, multi-physics and used inverse optimal design.

A one day RANS (Reynolds-Averaged Navier-Stokes) study for aircraft multidisciplinary analysis operated on 5 M points, 10 hours and 1 case. The next generations RANS + alpha, NSIII, should solve 30 M points, 10 hours, 10 cases. The same is true for combustion analysis. Thus NAL requested a machine that is 100 times faster than the actual -> 10 Tflop/s and 100 times larger user memory -> 5TByte. The requirements: 100 TFlop/s performance, 5 TB memory, 500 TB storage, 1 GB/s data transfer rate around NSIII. It should support resource continuity from NSII, standard and common, usability, Grid computing and Web computing. The operating system should be 64 bit Unix, transparent fast file system, one system image from the user.

The Central NS System CeNSS

NAL chose a Fujitsu Primepower HPC2500 with 14 compute cabinets, installed in 2002. Each consists of 128 way SMP with 256 GB shared memory. This leads to a peak performance of 665.6 GFlop/s per cabinet and thus to 9.3 TeraFlop/s total performance and 3.6 TB of aggregated memory. A node (128 CPUs) is a unit in terms of the OS and organised according to their roles, compute, service, and I/O nodes. Compared to the Numerical Wind Tunnel there is a factor of 33 in performance and 80 in terms of memory. Thus there is a performance improvement of 9300 times in 15 years.

Operating System and Parallel Environments

The Primepower runs Solaris 8, standard 64 bit Unix OS, which supports large systems, 128 CPUs/512 GB memory, has a high security level, various network capabilities and is installed on every node. The operation is still alive even when either one of the nodes is down, this is similar to a cluster model.

The high performance computing features are large page, which enhances memory access performance for huge programmes, harmonised scheduling, which ensures actual performance for parallelised programmes. The high-speed network file system is SRFS (Shared Rapid File System) with excellent data transfer performance by DTU and ensures file compatibility. Additionally there is a high-speed link to the visualisation system, GSN Link. GSN x 4stripes -> 500 MB/s sustained transfer rate and low overhead data transfer by ST protocol.

The Job Control Environment consists of a Batch system, the Network Queueing System (NQS) - substantial standard as a UNIX batch system, and resource management for jobs, guaranteed job-executing time by assigning system resources. There is a NAL intelligent job scheduler with job management which controls the number of jobs executed simultaneously, the Quick turn-around with better response and better algorithm. There is a wide variety of usage with interactive TSS and real time visualisation.

The Programme Developing Environment contains the Compilers F77, F90, F95, C, C++, parallel coding support with MPI, XPFortran (Multi processes) and automatic parallelisation, OpenMP (Multi threads). The tools are the Parallelnavi Workbench with an integrated development environment for parallel programming, a source code debugger and a Profiler. The mathematical library contains SSL II, C-SSL II, BLAS, LAPACK and ScaLAPACK.

Yuichi Matsuo mentioned some applications like Nastran, structural analysis, VisLink, real time visualisation and Public domain software GNU or CFD, UPACS. The transition from vector to scalar was successfully completed, the results for challenging problems are emerging, and the flexible combinations of process parallel and thread parallel makes the system effective.

He summarised:"NAL has been promoting supercomputer systems development and operation as the leading institution in Japan. The Numerical Simulator (NS) is one of the major facilities at NAL and has been serving for the Computational Fluid Dynamics (CFD) research. Since 1993 until 2002 the "Numerical Wind Tunnel" has been the main computing engine of the NS and was ranked as No.1 on the top500 list in 1993. Last October a new "Central Numerical Simulation System" (CeNSS) was installed. CeNSS is a SMP-cluster type distributed parallel system consisting of about 1800 scalar processors with 9.3Tflop/s peak performance and 3.6TB user memory. We are entering into a new era utilising the terascale computing power."