DFT requirements for leadership-class computers N. Schunck Department of Physics Astronomy, University of Tennessee, Knoxville, TN-37996, USA Physics Division, Oak Ridge National Laboratory, Oak Ridge, TN-37831, USA The 3rd LACM-EFES-JUSTIPEN Workshop JIHIR, Oak Ridge National Laboratory, February 23-25, 2009 A. Baran, J. Dobaczewski, J. McDonnell, J. Moré, W. Nazarewicz, N. Nikolov, H. H. Nam, J. Pei, J. Sarich, J. Sheikh, A. Staszczak, M. V. Stoitsov, S. Wild
Nuclear DFT: Why supercomputing? 1 Why super-computers: Large-scale problems (LACM): fission, shape coexistence, time-dependent problems Systematic restoration of broken symmetries and correlations “made easy” (QRPA, GCM?, etc.) Optimization of extended functionals on larger sets of experimental data DFT: A global theory Supercomputers: DFT at full power… Ground-state of even nucleus can be computed in a matter of minutes on a standard laptop: why bother with supercomputing? Principle: average out individual degrees of freedom Treatment of correlations ? Current lack of quantitative predictions at the ~100 keV level Extrapolability ? “No limit” theory: from light nuclei to the physics of neutron stars Rich physics Fast and reliable
Classes of DFT Solvers 2 1D2D3D r-space 1 mn, 1 core (HFBRAD) 5 hours,70 cores (HFBAX) - HO basis- 2 mn, 1 core (HFBTHO) 5 hours, 1 core (HFODD) Computational package used and developed at ORNL and estimate of the resources needed for a standard HFB calculation Coordinate-space: direct integration of the HFB equations Accurate: provide « exact » result Slow and CPU/memory intensive for 2D-3D geometries Configuration space: expansion of the solutions on a basis (usually HO) Fast and amenable to beyond mean-field extensions Truncation effects: source of divergences/renormalization issues Wrong asymptotic unless different bases are used (WS, PTG, Gamow, etc.) Non-linear integro-differential fixed point problem
Recent physics achievements 3 Even-even, odd-even and odd-odd mass tables Nuclear fission Systematics of odd-proton states in odd nuclei Cf. Talks by M. Stoitsov, S. Wild and J. Moré Online resources:
Petascale and beyond 4 Hardware constraints (see R. Lusk and J. Vary’s talks): Many cores (100,000+) stacked into sockets - Currently 4 cores/socket, evolution toward 8 cores/socket and more Small-memory per core (shared memory per socket) Short, crash-prone, expensive runtime Consequences on the architecture of DFT solvers: Optimize time of one HFB calculation: reduce number of iterations, use symmetries smartly by improving/interfacing codes, parallelization, etc. Work on parallel wrapper: load balancing, checkpoints, error control mechanisms, etc.
Optimization - Interface HFBTHO/HFODD Restarting HFODD from HFB-THO means: –Tremendous gain in time of calculation –Accrued numerical stability –Taking advantage of existing mass tables Procedure: –Coordinate + phase transformation (both unitary) –Modify HFODD to restart from HFB matrix elements instead of density fields on Gauss-Hermite mesh 5 Interface fulling working for spherical HO bases (precision of restart at ) Memory issue for deformed bases HFB-THO: Axial Cylindrical coordinates Time-reversal symmetry j-block diagonalization HFODD: symmetry- unrestricted Cartesian coordinates Y-simplex eigenbasis No time-reversal symmetry Full diagonalization
6 Optimization – HFODD Profiling Broyden routine: storage of N Broyden fields on 3D Gauss- Hermite mesh Temporary array allocation for HFB matrix diagonalization neutronsprotons Calculations by J. McDonnell Safe limit memory/core on Jaguar/Franklin
7 Optimization – HFODD Parallelization M M Two levels of parallelism handled by simple MPI group structure –Nuclear configuration (Z, N, interaction, {Q λμ }, etc.) –HFB solver Standard PBLAS and ScaLAPACK libraries for distributed linear algebra Natural splitting of the HFB matrix (OpenMP): perhaps not scalable enough Splitting: –HFB matrix into N blocks –Eigenfunctions conserve the same N-blocks splitting –Densities must be re-constructed piecewise Challenges –Identify self-contained set of all matrices required for one iteration –Handling of conserved symmetries: give different block structure –Identify and replace all BLAS calls by PBLAS equivalents M M
Optimization - Finite-size spin instabilities 8 Response of the nucleus to a perturbation with finite momentum q studied in the RPA theory Channels: scalar-isoscalar, scalar- isovector, vector-isoscalar, vector- isovector, etc. Modern Skyrme functionals are highly- instable with respect to finite-size spin perturbations ! Convergence of the HFB calculation of 100 blocked states in Ba Region of instability T. Lesinski et al, Phys. Rev. C 74, (2006) D. Davesne et al, arXiv: (2009) Warning for next generation of functionals: stability must be assessed !
Work in progress - Fission 9 Example of challenges for next generation DFT: microscopic description of nuclear fission Degrees of freedom at the HFB level: deformation, temperature Potential energy surfaces depend critically on interaction/functional and pairing correlations Computational tools – Augmented Lagrangian Method – Broyden Method Precision tools – Large bases – Benchmarks Distributed computing tools – MPI wrapper – Load balancing – Efficient, independent, constraint calculations Static HFB pre-requisites
DFT Computing Infrastructure 10 Interfacing codes Parallelize solver Load balancing
11 Deliverables Year 2-3 Have a DFT package combining HFB- THO and HFODD available for large- scale calculations Optimize full diagonalization of “large” (4,000 4,000) matrices in HFODD – Take advantage of N-core architecture – Increase speed for large bases (fission, heavy nuclei) – Overcome current memory limitations Optimize Broyden method (Cf. Jorge’s talk) to improve stability/convergence Papers on odd nuclei: 1.Methodology and Theoretical Models 2.Systematic and comparison with experiment Workplan Year 2-3Current Status Done (for spherical bases) - large-scale calculations up to 14,112 cores (2 hours) Well on target – Parallelization of the HFODD core (PBLAS, ScaLAPACK) – Will solve issues related to speed, memory and precision – Change of iteration cycle: updating HFB matrix elements instead of fields Done - Numerical instabilities of large-scale calculations can be tracked down to physical instabilities built-in current functionals (see Mario’s talk) Delayed by problem of instabilities – Paper 1 ready to be published – Paper 2 in preparation – Additional Paper 3 on finite-size spin instabilities in preparation
Work Plan (Year 4) 12 Physics – Optimization of DME-based functionals: genetic algorithm + Argonne optimizer (cf Mario’s talk) – Applications of DME functionals: UNEDF-1 Computing – Implement DME functionals in HFODD (study of time-odd channels) – Complete version 1.0 of parallel HFODD core Demonstrate efficiency and scalability of the code First applications: N-dimensional potential energy surface, fission pathways – Improve parallel interface to HFODD: Optimistic: it should be a good application of ADLB (“moderately long to long” work units of 1-2 hours, little communication). Realistic: remove the master and have him work like a slave (French revolution spirit) – Replace sequential I/O by parallel I/O for HFODD records (used as checkpoints) Remaining of the year New version of HFODD: HFBTHO interface, shell correction, finite-temperature, Augmented Lagrangian Method, matrix elements mixing, parallel interface, etc. 2 papers on odd nuclei and 1 on spin instabilities in preparation
Slide 14 Nuclear Structure and Nuclear Interactions Forefront Questions in Nuclear Science and the Role of High Performance Computing January 26-28, 2009 · Washington, D.C. December 10, 2008 Microscopic Description of Nuclear Fission Scientific and computational challenges Describe dynamics with novel energy functionals and ab initio methods 1)adiabatic approach 2)non-adiabatic/early stochastic 3)full time-dependent dynamics Develop ultra-scale techniques for the description of fission Build a spectroscopic precision nuclear energy density functional Perform constrained minimization on a multi- dimensional potential energy surface Find full spectrum of dense millions-sized matrices Predict half-lives, mass and kinetic energy distribution of fission fragments and fission cross-sections Analyze the fission process through the visualization of time evolution Develop scalable application software for time-dependent many-body dynamics Societal Impact Nuclear Energy programs Threat reduction NNSA Stockpile Stewardship Program Time-dependent many-body dynamics Low-energy heavy-ion collisions and nucleon- and photon-induced reactions Neutron star quakes Vortex dynamics in quantum super-fluids Summary of research direction Expected Scientific and Computational Outcomes Potential impact on Nuclear Science Our Holy Grail…