crinavar / trueke

multi-GPU Exchange Monte Carlo for 3D random field Ising model

trueke: multi-GPU Monte Carlo for the 3D RFIM

(1) Hardware requirements:

• A CUDA capable GPU, we recommend Kepler+
• 4GB of RAM
• Multi-core X86_64 CPU

(2) Software requirements:

• Linux OS
• GCC.
• Nvidia CUDA runtime library
• Nvidia nvcc compiler
• [optional] Nvidia nvlm (to query the device).
• OpenMP 4.0 Implementation

(3) Check Makefile and CUDA instalation.

Make sure the bin, inc and lib paths to the corresponding ones.

(4) compilation (edit Makefile if necessary)

• make clean
• make

(5) how to run

./bin/trueke -l <L> <R> -t <T> <dT> -a <tri> <ins> <pts> <ms> -h <h> -s <pts> <mz> <eq> <ms> <meas> <per> -br <b> <r> -z <seed> -g <x>

(6) parameters

• Lattice (-l): size <L>, <R> replicas.
• Temperature (-t): high-temp <T>, delta <dT>.
• Adaptive (-a): <tri> trials, <ins> inserts/trial, <pts> exchange steps, <ms> sweeps per exchange.
• External Field (-h): magnetic field strength <h> (tipically, 0 < h < 3).
• Simulation (-s): <pts> ex, meas at <mz>, equil <eq> sweeps, <ms> sweeps/ex, <per> ex/<meas>
• Repetition (-br): <b> blocks of <ms>, <r> disorder realizations.
• Seed (-z): use <seed> as base seed for the PRNGs.
• Multi-GPU (-g): use <x> GPUs.

(4) Example execution using two GPUs

• ./bin/trueke -l 64 11 -t 4.7 0.1 -a 58 2 2000 10 -h 1.0 -s 5000 3000 100 5 1 1 -br 1 2000 -z 7919 -g 2

trueke's Issues

questions about the kernels

I have some questions about your kernels.
1 Do you mind providing a CPU version (serial/openmp) of the metropolis kernel for reference and verification ?
2 When the maximum block size is 256 on a GPU, what are the values of BX, BY, and BZ ?

/* warning: changing the block dimensions can lead to undefined behavior. */
#define WARPSIZE 32
#define BX  32
#define BY  8
#define  BZ  4