Supercomputers and Their Performance in Computational Fluid Dynamics (Notes on Numerical Fluid Mechanics and Multidisciplinary Design, 37, Band 37)

Supercomputer technologies have evolved rapidly since the first commercial-based supercomputer, CRAY-1 was introduced in 1976. In early 1980's three Japanese super­ computers appeared, and Cray Research delivered the X-MP series. These machines including the later-announced CRAY-2 and NEC SX series created one generation of supercomputers, and the market was spread dramatically. The peak performance was higher than 1 GFLOPS and the compiler improvement was remarkable. There appeared many articles and books that described their architecture and their performance on The late 1980's saw a new generation of supercomputers. several benchmark problems. Following CRAY Y-MP and Hitachi S-820 delivered in 1988, NEC announced SX-3 and Fujitsu announced the VP2000 series in 1990. In addition, Cray Research announced the Y-MP C-90 late in 1991. The peak performance of these machines reached several to a few ten's GFLOPS. The hardware characteristics of these machines are known, but their practical performance has not been well documented so far. Computational Fluid Dynamics (CFD) is one of the important research fields that have been progressing with the growth of supercomputers. Today's fluid dynamic re­ search cannot be discussed without supercomputers and since CFD is one of the im­ portant users of supercomputers, future development of supercomputers has to take the requirements of CFD into account. There are many benchmark reports available today. However, they mostly use so called kernels. For fluid dynamics researchers, benchmark test on real fluid dynamic codes are necessary.

I.: Cray Y-MP C90 Supercomputer.- 1. Introducing The Cray Y-MP C90 Supercomputer.- 2. Redefining High-Performance Computing.- 3. Bridging the Gap Between Potential and Productivity.- 4. Protecting Your High-End Supercomputing Investments.- 5. The Best Overall Supercomputing Solutions.- 6. New Technologies Maximize System Availability.- 7. Physical Description.- 8. Cray Y-MP C90 Highlights.- 9. The Most Powerful I/O Technology Available.- 10. Input/Output Highlights.- 11. Advanced SSD Technology.- 12. SSD Highlights.- 13. Disk Drives.- 14. Software.- 14.1 Performance Oriented, Feature-Rich Software.- 14.2 UNICOS Operating System.- 14.3 UNICOS Highlights.- 14.4 Compilers.- 14.5 Autotasking.- 14.6 UNICOS Storage System.- 14.7 Applications.- 14.8 The Power of Visualization.- 15. Network Supercomputing.- 15.1 Delivering Supercomputing Power to Your Desktop.- 16. Supportability.- 16.1 Maximized System Availability.- 17. The Cray Y-MP C90 Supercomputer, Nothing Else Comes Close.- II. Fujitsu VP2000 Series Supercomputer.- 1. Introduction.- 2. Architecture.- 2.1 Scalar Unit (SU).- 2.2 Vector Unit (VU).- 2.3 Main Storage Unit (MSU).- 2.4 System Storage Unit (SSU).- 2.5 Channel Processor (CHP).- 2.5.1 High-speed optical channel.- 2.5.2 HIPPI channel.- 3. Hardware Implementation.- 3.1 Vector Pipelines.- 3.2 Parallel Processing.- 3.3 Advanced Scalar Operation.- 3.4 Other Features for High Speed Processing.- 4. Multiprocessor System.- 4.1 Dual Scalar Processor (DSP).- 4.2 Quadruple Scalar Processor (QSP).- 5. Hardware Technology.- 5.1 Advanced LSIs.- 5.2 High Density Packaging.- 5.3 Cooling Technology.- 6. MSP System.- 6.1 System Storage Usage.- 6.1.1 High speed large scale virtual I/O.- 6.1.2 High speed swapping.- 6.2 Support of DSP/QSP.- 6.3 Virtual Machine.- 6.4 TCP/IP Support.- 7. Unix System.- 7.1 Optimization of Vector Processes.- 7.2 High-Speed I/O Access.- 7.3 Effective Resource Management.- 7.4 High-Speed Swapping.- 8. Language Processing System.- 8.1 Optimization.- 8.1.1 Parallel pipeline scheduling (PPS).- 8.1.2 Loop unrolling.- 8.2 Parallelization.- 8.2.1 Automatic parallelization.- 8.2.2 Parallelism description.- 9. Performance.- 10. Conclusion.- 11. References.- III. Hitachi S-820 Supercomputer System.- 1. Introduction.- 2. Architecture and System Organization.- 2.1 Overview.- 2.2 Extended Storage.- 2.3 Vector Register.- 2.4 Vector Instruction Set.- 3. Logic Structure.- 3.1 Overview.- 3.2 Vector Execution Control.- 3.2.1 Parallel construction.- 3.2.2 Elementwise parallel processing.- 3.3 Storage Control.- 4. Hardware Technology.- 5. Software.- 6. Performance.- 7. Conclusion.- 8. References.- IV. NEC SX-3 Supercomputer System.- 1. Introduction.- 2. System Configuration.- 3. Processor Configuration and Architecture.- 4. The Super-Ux Operating System.- 5. Fortran and Tools.- 6. Performance Results.- 7. Conclusion.- 8. References.- V. Trends in Vector and Parallel Supercomputer Architectures.- 1. Introduction.- 2. The Supercomputer CPU: An Overview.- 3. A Summary of Supercomputer Hardware Characteristics.- 4. Parallel Vector Computation, and Latency in Design.- 5. A Study of Vector Start-Up Time.- 6. Parallel Computation.- 7. Risc Architectures.- 8. Conclusion.- 9. References.- VI. Navier-Stokes Benchmark Tests.- 1. Introduction.- 2. Benchmark Test Features.- 3. Benchmark Test Result ― 1.- 4. Benchmark Test Result ― 2.- 5. Final Remarks On Both Benchmark Tests.- 5.1 Assessment of the Result.- 5.2 CFD View Point.- 6. Cray Y-MP C-90 Benchmark Report.- 7. Future Requirements.- 8. Final Remarks.- 9. Acknowledgment.- 10. References.- VII. Vectorization and Parallelization Techniques for Modern Supercomputers.- 1. Introduction.- 2. Basic Aspects of Vector and Parallel Processing.- 2.1 Vector Architectures and Vector Processing.- 2.2 Parallel Architectures and Parallel Processing.- 2.3 Shared-Memory Systems.- 2.4 Distributed-Memory Systems.- 3. Vectorization and Parallelization of Algorithms.- 3.1 Vectorization.- 3.2 Parallelization.- 3.3 Example: Restructuring of the SOR-Poisson Solver.- 3.4 Example: Vectorization of Sparse Matrix Vector Products.- 3.5 Parallelization of SOR for Shared-Memory Systems.- 3.6 Parallelization of SOR for Distributed-Memory Systems.- 3.7 Example: Numerical Grid Generation.- 4. Concluding Remarks.- VIII.: UHSNWT Initiative at National Aerospace Laboratory.- 1. Background of Numerical Wind Tunnel.- 1.1 Present Situation of CFD.- 1.2 From Ultra High Speed Supercomputer to Ultra High Speed Numerical Wind Tunnel.- 2. Demands in the System Manager’s Eyes.- 2.1 Costs.- 2.2 Reliability.- 3. The Uhsnwt Initiative.- 3.1 Starting Point.- 3.2 Hierarchical Structure of the UHSNWT Memory.- 3.3 Required Performance ― From the Manager’s Viewpoints.- 3.4 Configuration of PE.- 3.4.1 Speed-up of PE.- 3.4.1.1 Pipelined vector computers with large VR.- 3.4.1.2 VTAP simulation.- 3.4.2 VTAP simulation results.- 3.4.3 PE model and its feasibility.- 3.4.4 PE models.- 3.4.5 Analysis of VTAP simulation.- 3.4.6 LSI chips for PE.- 3.5 Configuration of Main Memory.- 3.5.1 Realization of target main memory capacity.- 3.5 2 Affinity with CFD programs.- 4. Overall Hardware Configuration of the Uhsnwt.- 4.1 Summary.- 4.2 Reliability of the UHSNWT.- 4.3 Overall Performance.- 4.4 Feasibility of UHSNWT Meeting Requirement (R2).- 5. Concluding Remarks.- 6. References.- IX. Addresses of Contributors.