Scribd
Upload
198 views17 pages

Lammps On Gpus: A Tutorial

This document provides a tutorial on running molecular dynamics simulations using LAMMPS on GPUs. It discusses why GPUs are useful for scientific computing due to their large number of cores and high memory bandwidth. It then summarizes the ongoing efforts to port LAMMPS to GPUs and the capabilities that are currently available like Lennard-Jones and Gay-Berne potentials. The rest of the document outlines the 9 step process to run LAMMPS on a GPU, including checking your GPU, installing CUDA, editing Makefiles, compiling the GPU library, adding GPU packages to LAMMPS, modifying input scripts, and running a sample simulation. Speedups achieved will depend on factors like the CPU, GPU,

Uploaded by

sachu92
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
198 views17 pages

Lammps On Gpus: A Tutorial

This document provides a tutorial on running molecular dynamics simulations using LAMMPS on GPUs. It discusses why GPUs are useful for scientific computing due to their large number of cores and high memory bandwidth. It then summarizes the ongoing efforts to port LAMMPS to GPUs and the capabilities that are currently available like Lennard-Jones and Gay-Berne potentials. The rest of the document outlines the 9 step process to run LAMMPS on a GPU, including checking your GPU, installing CUDA, editing Makefiles, compiling the GPU library, adding GPU packages to LAMMPS, modifying input scripts, and running a sample simulation. Speedups achieved will depend on factors like the CPU, GPU,

Uploaded by

sachu92
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

LAMMPS on GPUs

A Tutorial
W. Michael Brown, Peng Wang, Paul S. Crozier, Steve Plimpton

Wednesday, February 24, 2010

Why run on GPUs?


Technology paid for by
gamers, but impact to scientific computing is now well-recognized (electrical) solution for data parallelism 240+ cores on a GPU High memory bandwidth

Cheap, low-power

Porting LAMMPS to GPUs


Still largely a research effort
Marc Adams (Nvidia) Pratul Agarwal (ORNL) Sarah Anderson (Cray) Mike Brown (Sandia) Paul Crozier (Sandia) Massimiliano Fatica (Nvidia) Scott Hampton (ORNL) Ricky Kendall (ORNL) Hyesoon Kim (Ga Tech) Axel Kohlmeyer (Temple) Doug Kothe (ORNL) Scott LeGrand (Nvidia) Ben Levine (Temple) Christian Mueller (UTI Germany) Steve Plimpton (Sandia) Duncan Poole (Nvidia) Steve Poole (ORNL) Jason Sanchez (RPI) Arnold Tharrington (ORNL) John Turner (ORNL) Peng Wang (Nvidia) Lars Winterfeld (UTI Germany) Andrew Zonenberg (RPI)

Currently Available in Main LAMMPS


Lennard-Jones
Force/Neighbor

Gay-Berne Potential
Force

More capabilities
soon

How to Run LAMMPS on Your GPU

1. Do you have a GPU?


For single precision
Currently need a CUDA-enabled GPU with compute
capability >= 1.1

For double precision


Currently need a CUDA-enabled GPU with compute
capability >= 1.3
Windows: Device Manager Apple: Apple Menu-> About this Mac -> More Info -> Graphics/Displays Linux: nvidia_settings or /sbin/lspci | grep nVidia List of CUDA-enabled GPUs here: http://www.nvidia.com/object/cuda_gpus.html Can use device query to get compute capability; more later

2. Do you have CUDA?


http://developer.nvidia.com/object/
cuda_2_3_downloads.html Need driver and toolkit only Need to have the nvcc compiler in your path Pay attention to 32- or 64-bit
No 64-bit on apple!
set path = ( $path /usr/local/cuda/bin ) setenv LD_LIBRARY_PATH /usr/local/cuda/lib/ or set path = ( $path /usr/local/cuda/bin ) setenv LD_LIBRARY_PATH /usr/local/cuda/lib64/

3. Edit LAMMPS GPU Makefile

set LROOT = /home/wmbrown/lammps-20Feb10 cd $LROOT/lib/gpu emacs Makefile.nvidia

3. Edit LAMMPS GPU Makefile (2)


BIN_DIR = . OBJ_DIR = . AR = ar CUDA_CPP = nvcc -I/usr/local/cuda/include -DUNIX -O3 -Xptxas -v -use_fast_math CUDA_ARCH = -arch=sm_13 CUDA_PREC = -D_SINGLE_SINGLE CUDA_LINK = -L/usr/local/cuda/lib64 -lcudart $(CUDA_LIB)

For compute capability >= 1.3 can also use:


CUDA_PREC = -D_SINGLE_DOUBLE # Double precision accumulation or CUDA_PREC = -D_DOUBLE_DOUBLE # Double precision everything

For Apple, must compile 32-bit


CUDA_ARCH = -arch=sm_13 m32 CUDA_LINK = -L/usr/local/cuda/lib -lcudart $(CUDA_LIB)

For compiler >= g++ 4.4 on Linux


CUDA_ARCH = -arch=sm_13 --compiler-bindir=/usr/bin/gcc-4.3

4. Make LAMMPS GPU lib


make f Makefile.nvidia ./nvc_get_devices
Device 0: "GeForce GTX 295" Revision number: Total amount of global memory: Number of multiprocessors: Number of cores: Total amount of constant memory: Total amount of shared memory per block: Total number of registers available per block: Warp size: Maximum number of threads per block: Maximum sizes of each dimension of a block: Maximum sizes of each dimension of a grid: Maximum memory pitch: Texture alignment: Clock rate: Concurrent copy and execution: Device 1: "Tesla C1060" 1.3 0.87 GB 30 240 65536 bytes 16384 bytes 16384 32 512 512 x 512 x 64 65535 x 65535 x 1 262144 bytes 256 bytes 1.24 GHz Yes

5. Edit LAMMPS Makefile as Necessary


cd $LROOT/src emacs ./MAKE/Makefile.linux If you are not 64-bit (or Apple) gpu_SYSPATH = -L/usr/local/cuda/lib

If you are using Apple, compile LAMMPS 32-bit to link with GPU library CC = LINK = g++ -m32 g++ -m32

make clean

6. Add GPU Package to LAMMPS


cd $LROOT/src make yes-asphere make yes-gpu make linux

7. Modify your input script


cd $LROOT/bench emacs in.lj Must add newton off to beginning of script and /gpu to a supported pair_style
newton off ... pair_style lj/cut/gpu one/node 0 2.5

GPU Selection Keyword

GPU ID

7. Modify your input script (2)


GPU Selection Keyword
one/node - single compute "node, which may have
multiple cores and/or GPUs. GpuID should be set to the ID of the (first) GPU you wish to use with LAMMPS one/gpu - multiple compute "nodes with one GPU per node. GpuID should be set to the ID of the GPU. multi/gpu - multiple compute "nodes" on your system with multiple GPUs. GpuID should be set to the number of GPUs per node

8. Run your input script


Number of procs = number of gpus you want
mpirun np 3 lmp_linux < in.lj
-------------------------------------------------------------------------- Using GPGPU acceleration for LJ-Cut: -------------------------------------------------------------------------GPU 1: Tesla C1060, 240 cores, 4 GB, 1.3 GHZ GPU 2: Tesla C1060, 240 cores, 4 GB, 1.3 GHZ GPU 3: GeForce GTX 295, 240 cores, 0.87 GB, 1.2 GHZ ---------------------------------------------------------------------------------------------------------------------------------------------GPU Time Stamps: --------------------------------------------------------------------Atom copy: 0.07111 s. Neighbor copy: 0.0004615 s. LJ calc: 0.1702 s. Answer copy: 0 s. ---------------------------------------------------------------------

9. Speed-ups
Depends on

Your CPU Your GPU Number of Particles Cutoff

More talks showing the GPU acceleration in


LAMMPS to come

Questions

You might also like