Can the Earth Simulator change the way humans think Tetsuya Sato - PowerPoint PPT Presentation

Can the Earth Simulator change the way humans think Tetsuya Sato The Earth Simulator Center Japan Marine Science and Technology Center Yokohama, Japan

Outline • What is the Earth Simulator • Can Computer Simulation open up a new science field • Can global prediction be made accurately • Can paradigm be changed

The Earth Simulator (ES) is a Vector-parallel supercomputer � Development of ES started in 1997 in order to make a comprehensive understanding of global changes such as global warming. � Its construction completed at the end of February, 2002 and the practical operation started from March 1, 2002 � 35.86Tflops (87.5% of the peak performance) is already achieved in the Linpack benchmark. � 2 ６ ． ５ ８ Tflops was obtained by a global atmospheric circulation code.

Location of Earth Simulator Facilities Tokyo Earth Simulator Site

Earth Simulator Building

Artist’s View of the Earth Simulator Cartridge Tape Library System Magnetic Disk System PN cabinets(320) IN cabinets(65) 65m (71yd) Double Floor for Cables Power Supply System Air Conditioning System Seismic Isolation System 50m (55yd)

Features of Earth Simulator Building Steel-frame construction, Two-story 65m x 50m (71yd x 55yd) 17m (19yd) height Lightning conductor isolated from building Shielding by steel plates from electromagnetic waves Air return duct Rooms surrounding with steel nets Air conditioning system Power supply system Seismic isolation system ( Insulation by 11 layered rubbers )

Another Photo of Earth Simulator

Arithmetic Processor Package Heat sink LSI Arithmetic Processor (AP) Clock Connector Build-up PCB Power Connector MMU Interface Connectors ( 115mm x 139mm )

Configuration of the Earth Simulator • Peak performance/AP : 8Gflops • Total number of APs : 5120 • Peak performance/PN : 64Gflops • Total number of PNs : 640 • Shared memory/PN : 16GB • Total peak performance: 40Tflops • Total main memory : 10TB Interconnection Network (full crossbar switch) Shared Memory Shared Memory Shared Memory 16GB 16GB 16GB Arithmetic Processor #0 Arithmetic Processor #1 Arithmetic Processor #7 Arithmetic Processor #0 Arithmetic Processor #1 Arithmetic Processor #7 Arithmetic Processor #0 Arithmetic Processor #1 Arithmetic Processor #7 Processor Node #0 Processor Node #1 Processor Node #639

XCT #0 PN #0 XCT #1 PN #1 PN #2 PN-IN Electric Cables : 640 x 130 = 83,200 XSW #0 PN #3 Connection between Cabinets XSW #1 XSW #2 PN #4 XSW #3 PN #5 XSW #4 XSW #5 320 Cabinets XSW #6 640 PNs XSW #7 64 Cabinets 128 XSWs PN #636 PN #637 XSW #126 PN #638 XSW #127 PN #639

Two XSWs in Cabinet Cooling Fan Diagnostic Processor (DGP) AC-DC Converter 2 0 0 c m Serial/Parallel Converter XSW DC-DC Converter PN-IN Cable Power Supply Unit 1 2 0 c m 1 3 0 c m

Two Processor Nodes in Cabinet Main Memory Unit (MMU) Arithmetic Processor(AP) Remote access Control Unit(RCU) 140cm Cooling Fan DC-DC Converter 200cm AC-DC Converter 100cm Diagnostic Processor (DGP) I/O Processor (IOP) Serial/Parallel Power Supply Unit PN-IN Cable Converter

PN Cabinet IN Cabinet 200cm 140cm 120cm 130cm 100cm

Electric Cables Connecting Cabinets

Comparison of PN Size NEC SX-4 (1 node) Earth Simulator Peak Performance : 64Gflops Peak Performance: 64Gflops Electric Power : about 90kVA Electric Power : about 8kVA Air Cooling Air Cooling 100cm about 7m 70cm about 6m

Software Environment � Operating System � UNIX-based system (Enhanced version of NEC SUPER-UX) � Parallel file system ( MPI-IO, HPF ) � Programming Environment � Parallel programming environment ({Fortran90,C}+MPI2, HPF2) � Program analyzer + Tuning tools � Parallel debugger � Flexible job scheduler � Earth Simulator own specification � Running on the SCCS � Job assignment to PNs with file loading to appropriate system disks The software has a good scalability up to 640 nodes.

T1279L96

Sustained Performance Sustained Performance AFES (T1279L96) is based on the spectral method. CPU/Node ELAPSE Tflops Ratio Total CPU Node ( sec ) Peak (%) Sustained 80 80 1 238.04 0.52 0.64 81.1 160 160 1 119.26 1.04 1.28 81.0 320 320 1 60.52 2.04 2.56 79.8 640 80 8 32.06 3.86 5.12 75. 3 1280 160 8 16.24 7.61 10.24 74.3 2560 320 8 8.52 14.50 20.48 70.8 5120 640 8 40.96 4.65 26.58 64.9 26.58 64.9 Number of time integration steps : 10

Sustained Performance 40.00 35.00 30.00 25.00 Tflops 20.00 15.00 10.00 5.00 0.00 0 640 1280 1920 2560 3200 3840 4480 5120 Number of CPUs Sustained Peak

Summary � The world’s fastest supercomputer, the Earth Simulator, is successfully completed with 40Tflops theoretical peak performance. � 35.86Tflops sustained performance is obtained in the Linpack benchmark. � Global atmospheric and oceanic circulation codes were optimized and executed on the ES and the atmospheric code achieved an excellent performance of 26.58Tflops. � The ES went into preparatory operation in March of 2002. The regular operation will come soon.

Can Computer Simulation open up a new science field