R. Ryne, NUG mtg: Page 1 High Energy Physics Greenbook Presentation Robert D. Ryne Lawrence Berkeley National Laboratory NERSC User Group Meeting June 25, 2005
R. Ryne, NUG mtg: Page 2 Outline Lattice QCD Accelerator Physics Astrophysics (see D. Olson’s presentation)
R. Ryne, NUG mtg: Page 3 Lattice QCD
R. Ryne, NUG mtg: Page 4 Goals Determine a number of basic parameters of the Standard Model Make precise tests of the Standard Model Obtain a quantitative understanding of the physical phenomena controlled by the strong interactions
R. Ryne, NUG mtg: Page 5 Impact on determination of CKM matrix Improvements in lattice errors obtained w/ computers sustaining 0.6, 6, and 60 Tflops for one year
R. Ryne, NUG mtg: Page 6 Computing Needs: Approach Two pronged approach: Use of national supercomputer centers such as NERSC Build dedicated computers using special purpose hardware for QCD - QCDOC - Optimized clusters Special purpose hardware is used to perform the majority of the lattice calculations Supercomputer centers used for a combination of lattice calculations and data analysis
R. Ryne, NUG mtg: Page 7 Computational Issues Lattice calculations utilize a 4D grid Need highest possible single processor performance Communication is nearest-neighbor Don’t need large memory Do need high speed networks - International Lattice Data Grid formed to share computationally expensive data - Need to move ~1 petabyte in 24 hrs
R. Ryne, NUG mtg: Page 8 Lattice QCD Computational Roadmap Lattice community presently sustains Tflop/sec Has allowed determination of a limited number of key quantities to ~few percent accuracy Has allowed development & testing of new formulations of that will significantly improve accuracy of future calculations In next few years need to sustain Tflop/sec Calculate weak decay constants & form factors Determine phase diagram of high temp QCD, calculate EOS of quark-gluon plasma Obtain quantitative understanding of internal structure of strongly interacting particles Need to sustain ~ 1 petaflop/sec by end of decade
R. Ryne, NUG mtg: Page 9 Accelerator Physics
R. Ryne, NUG mtg: Page 10 Goals Large-scale modeling is essential for Improving/upgrading existing accelerators Designing next-generation accelerators Exploring/discovering new methods of acceleration - Laser/plasma based concepts
R. Ryne, NUG mtg: Page 11 Accelerator modeling is very diverse Many models Maxwell Vlasov/Poisson Vlasov/Maxwell Fokker-Planck Leonard-Weichart Single & multi-species Particle based codes Mesh-based codes: regular, irregular, AMR,… Combined particle/mesh codes Runs of various sizes (up to ~1000 PEs and beyond)
R. Ryne, NUG mtg: Page 12 Advanced Computing: An imperative to help assure success and best performance of a ~$20B investment SciDAC budget is < 0.02% of the this amount Small investiment in computing can have huge financial consequences
R. Ryne, NUG mtg: Page 13 Accelerator Modeling Roadmap Current resources: ~3M NERSC In the next few years, will need ~20M hrs/yr Design of proposed machines: Linear Collider, RIA, hadron machines (proton drivers, muon/neutrino systems, VLHC) Simulation of existing & near-term machines: LHC,RHIC, PEP-II, SNS Design of advanced concepts: 1 GeV stage, plasma afterburner Design of 4th generation light sources By the end of the decade will need ~60M hrs/yr Full scale electron-cloud Multi-slice, multi-IP, strong-strong beam-beam Interaction of space charge effects, wakefields, and machine nonlinearities in boosters and accumulator rings First principles Langevin modeling of electron cooling systems CSR effects with realistic boundary conditions Goal is end-to-end modeling of complete systems
R. Ryne, NUG mtg: Page 14 Algorithmic & Software Needs Continued close collaboration with ASCR- supported researchers is essential Linear solvers, eigensolvers, PDE solvers, meshing technologies, visualization Performance monitoring and enhancement, version control & build tools, multi-language support Multi-scale methods are becoming increasingly important We need robust, easy-to-use parallel programming environments & parallel scientific software libraries
R. Ryne, NUG mtg: Page 15 Parallel Optimization promises to be well suited for design problems on 10’s of thousands of processors Machine design always involves multiple runs Up to now the community has learned how to run large problems on ~thousand processors In the future, it will be desirable to run multiple ~1000 processor runs in a single optimization step Will allow scaling up to 10’s of thousands of processors for machine design problems NOTE: not all problems are design problems. Fast interprocessor communication is needed for the very largest “single point” runs.
R. Ryne, NUG mtg: Page 16 Diversity of accelerator modeling problems demands a mix of capacity & capability, and a mix of system parameters Some problems well suited to <=500 processors, but we typically need to run a large # of simulations Design studies, parameter scans Some problems demand large simulations (>=1000) procs) and involve regular, near-neighbor comm. Electromagnetic PIC Some problems demand large simulations and involve global, irregular communication Modeling geometrically complex electromagnetic structures
R. Ryne, NUG mtg: Page 17 THE END