Large-scale MD simulation of heterogeneous systems with ls1 mardyn

Transcription

1 Large-scale MD simulation of heterogeneous systems with ls1 mardyn M. T. Horsch, R. Srivastava, S. J. Werth, C. Niethammer, C. W. Glass, W. Eckhardt, A. Heinecke, N. Tchipev, H.-J. Bungartz, S. Eckelsbach, J. Vrabec and H. Hasse TU Kaiserslautern, Engineering TU München, Scientific Computing in Computer Science (SCCS) High Performance Computing Centre Stuttgart (HLRS) University of Paderborn, Thermodynamics and Energy Technology (ThEt) Frankfurt am Main, 23rd March 2015 ProcessNet International Workshop MolMod

2 Parallelization by volume decomposition Linked-cell data structure suitable for spatial domain decomposition: (non-blocking, overlapping MPI send/receive operations) large systems 1 : molecular dynamics 2

3 Parallelization by volume decomposition Linked-cell data structure suitable for spatial domain decomposition: (non-blocking, overlapping MPI send/receive operations) Methods for heterogeneous or fluctuating particle distributions: 3

4 Scale separation and long-range correction For planar interfaces: short range (explicit) long range (correction) Long-range correction from the density profile, following Janeček. f ra to f cu s diu Full evaluation of all pairwise interactions is too expensive instead, short-range interactions are evaluated for neighbours. 4

5 Scale separation and long-range correction For planar interfaces: short range (explicit) long range (correction) Long-range correction from the density profile, following Janeček. f ra to f cu s diu Angle-averaging expression for multi-site models, following Lustig. Full evaluation of all pairwise interactions is too expensive instead, short-range interactions are evaluated for neighbours. 5

6 Molecular simulation of fluids at interfaces Long-range correction from the density profile, following Janeček. Angle-averaging expression for multi-site models, following Lustig. Two-centre LJ fluid (2CLJ) surface tension / εσ -2 For planar interfaces: 1 nm Janeček-Lustig term no angle averaging no correction at all cutoff radius / σ For arbitrary geometries, e.g. the fast multipole method can be employed. 6

7 Molecular simulation of fluids at interfaces 2CLJQ models: 2 LJ centres Quadrupole Test of predictivity for interfacial properties Model validation and optimization 7

8 Molecular simulation of fluids at interfaces Adsorption (fluid-fluid and fluid-solid) Vapour-liquid surface tension Curved vapour-liquid interfaces Contact angle and contact line pinning LJTS T = 0.8 ε θpl = 90 8

9 MD simulation of nanofluidics 9

10 Scale bridging from nano- to microfluidics 10

11 Scaling of ls1 mardyn on hermit 11

12 Scaling of ls1 mardyn on hermit homogeneous cavitation CO2 (T = 280 K and ρ = 17.2 mol/l), 3CLJQ 25 million molecules on cores 12

13 Optimization of ls1 mardyn for SuperMUC SuperMUC (LRZ Garching): 3 PFLOPS Intel Xeon Sandy Bridge cluster. forces acting on molecules are only stored while the cell is inside the sliding window hyperthreaded sliding window Efficient vectorization: Optimization by hand, using advanced vector extensions (AVX). Conversion from array of structures (AoS) to structure of arrays (SoA). Discussed in detail by Nikola Tchipev tomorrow. 13

14 Large-scale MD simulations on SuperMUC speedup (relative to 128 cores) Scaling of ls1 mardyn examined on up to cores, i.e. the whole SuperMUC, by Wolfgang Eckhardt and Alexander Heinecke in homogeneous LJTS liquid with 4.8 billion molecules l idea s alin c s g tron o g st d e v r bs e r ling a c s ong number of cores 14

15 Large-scale MD simulations on SuperMUC speedup Up to N = molecules on SuperMUC weak scaling with 31.5 million molecules per core 2013 number of cores 15

16 Release of ls1 mardyn rele Fre ased e (BS Soft as w Dl i c e a re n se ) Free registration for ls1 mardyn at 16