- version: 1.0.0 (based on mVMC 2013.11.08 version)
- date: 2014/07/17
- contact: miniapp@riken.jp
The mVMC-mini package is based on mVMC simulation program. The primary purpose of mVMC-mini is to enable the performance study of mVMC on various platforms with less effort using compact installation procedure while representing a typical computational work load of mVMC. It contains the mVMC source program and the test data as described in the following sections.
Regarding the license condition to use this mVMC-mini package, there are several related files. Please refer to the "LICENSE" file included in the top directory of this package. Please also refer to the two "LICENSE" files each included in src/pfapack/ and src/sfmt/ subdirectories which are used as part of mVMC.
The mVMC-mini package "mVMC-mini-1.0.tar.gz" can be downloaded from the
repository.
http://fiber-miniapp.github.io/
The following software suite is necessary for installing the package.
- Prerequisite software:
- C compiler and Fortran compiler with OpenMP support
- BLAS and LAPACK library.
- BLACS and ScaLAPACK library.
- MPI library.
####example - K computer As a quick start, the installation example on K computer is shown first. The installation should be quite easy as below.
$ tar -zxf mVMC-mini-1.0.tar.gz
$ cd src
$ make Kei
When the installation completes successfully, the executable file is created as "vmc.out" in the src/ directory as below.
$ ls -go vmc.out
-rwxr-xr-x 1 463739 Jun 12 17:28 vmc.out
####example - Intel compiler and libraries The installation example on Intel software environment is shown next.
$ tar -zxf mVMC-mini-1.0.tar.gz
$ cd src
$ make intel
####General installation steps on other Linux platform
The installation of mVMC-mini on other Linux based platforms should be fairly simple as well. A typical installation step is explained below.
#####step 1.
Obtain the mVMC-mini package "mVMC-mini-1.0.tar.gz" from the repository. http://fiber-miniapp.github.io/ Extract its contents using tar command. This readme.txt (readme.md) file is included in the package.
$ tar -zxf mVMC-mini-1.0.tar.gz
$ ls
LICENSE job_middle makeDef result
README.md job_tiny readme.asis.utf8 src
The directories contain the following files.
- job_middle/ # medium size test job definition directory
- job_tiny/ # tiny size test job definition directory
- makeDef/ # contains Python script to produce the job definition file
- result/ # contains computed results on a reference platform (FX10)
- src/ # source directory, with following additional subdirectories
- pfapack/ # subdir. containing Pfaffian computation library
- sfmt/ # subdir. containing SIMD-oriented Fast Mersenne Twister
In src/ directory, there are "Makefile_*" for several platforms.
- Predefined platform : Makefile_${platform}
- Makefile_intel : for Intel compiler+MPI on Intel Xeon Linux cluster
- Makefile_fx10 : for Fujitsu compiler+MPI on FX10 and K computer
- Makefile_pgi : for PGI compiler + ACML + OpenMPI/pgi on Linux
If the installing platform is either intel or fx10, then set the value of platform as so. For example of intel, set the value as below.
$ platform=intel
Or for K computer and/or FX10 system, set the value as below.
$ platform=fx10
In either case, you can move to step 3 without taking step 2.
If the installing platform has PGI compiler, then use the file "Makefile_pgi" and modify the values of LAPACK and SCALAPACK path. The combination of PGI/ACML/ScaLAPACK and their locations are system dependent in combination with MPI choice. After the values are set for LAPACK and SCALAPACK, then set platform value as:
$ platform=pgi
and you can move to step 3 without taking step 2.
If the installing platform is NOT covered by either of above, create an appropriate "Makefile_${platform}" by taking the following step 2.
#####step 2.
When installing mVMC on other platform with different compiler and MPI combination, create "Makefile_${platform}" file accordingly. The value for ${platform} can be chosen anonymously, such as akb.
Edit the file and set the compiler names and their options with CC , FC , CFLAGS and FFLAGS variables. mVMC-mini requires the preinstalled LAPACK library and ScaLAPACK library whose locations are system dependent. Link option LIB should point to these LAPACK and SCALAPACK.
The values of all these variables should be set in "Makefile_${platform}" file. A skeleton file "Makefile_skeleton" is provided for convenience. It may be easier to copy and edit the skeleton file as shown below.
$ platform=akb
$ cd src
$ cp Makefile_intel Makefile_${platform}
$ vi Makefile_${platform} # edit CC, CFLAGS, etc.
#####step 3.
By now the makefile "Makefile_${platform}" must be ready in src/ directory.
Run make command in src/ directory. After make command finishes successfully, there should be an executable file named "vmc.out" in the directory.
$ make -f Makefile_${platform}
$ ls -go vmc.out
-rwxr-xr-x 1 463739 Jun 12 17:28 vmc.out
A couple of test directories are included in the package.
- job_middle/ Medium size test job definition directory. It takes 4 minutes on K computer 128 nodes (128MPIx8OpenMP)
- job_tiny/ Tiny size test job definition directory. It takes 12 seconds on Intel E5-2420. Mostly used for a quick debug.
There is a shell script file "job.sh" in each directory. The script can be run in foreground. It can be run as a batch job with appropriate batch directives added.
After the successful job execution, the output files named "zvo_*.dat" are saved under a directory. The name of the directory is defined in the input data file "multiDir.def". In the stdout, there will be an error message printed as > "Error: opt.init does not exist." which can be safely ignored.
For medium size test job, example batch job script files for K computer, FX10 and Intel cluster are provided as:
- "job-K.sh"
- "job-FX10.sh"
- "job-Intel.sh"
On K computer, the input/output files must be staged. Please note that the stgin/stgout-basedir and stgin/stgout directives in "job-K.sh" must be modified according to the installed path. The jobscript lines starting with:
#PJM --stgin-basedir
#PJM --stgout-basedir
will have to be changed to reflect the base directories and files accordingly.
The simulated model is 2D square Kondo lattice model (J/t=1.0, half-filling) and 20 steps of parameter optimization is carried out. The number of parameter optimization steps is defined by NSROptItrStep in the "zmodpara.def" file. The computing time should be proportional to this value, so it can be used to control the job elapse time.
The previously computed result files on a reference platform (FX10) are also saved under result/ directory for comparison. Verifying the numerical results can be best done by checking the values in "zvo_out_000.dat" file. Its first column shows the value of the energy expectation. For jobs with the same number of MPI processes, the computed values should roughly match.
mVMC adopts two different parallel processing approaches for its computing phases. For the parameter optimization phase, it calls ScaLAPACK using all the MPI processes. For the Monte Carlo computation phase, it divides the MPI processes into NSplitSize groups specified in zmodpara.def file, and each group produces NVMCSample Monte Carlo samples and computes physical variables. The computed values from all the processes are gathered to calculate the expectation. If a job is configured to run with Nmpi processes, the total number of the Monte Carlo samples will be (Nmpi/NSplitSize)*NVMCSample . Nmpi is the number of MPIs, which is not included in the zmodpara.def file, but is declared as mpirun command option.
The default setting will effectively define weak scaling test variations. Copy the directory job_middle to another one, and change the number of MPIs to run weak scaling test.
The parameters in the default job_middle zmodpara.def includes NSplitSize=4 and NVMCSample=192, so if a job is run using 4 MPIs, the number of Monte Carlo sampling will be 192, and if a job is run using 8 MPIs, the number will be 384, etc.
For strong scaling tests, the number of Monte Carlo samples should be kept the same across the tests, which can be done by changing the number of groups NSplitSize according to Nmpi.
A separate Python script named "makeDef_large.py" under makeDef/ directory is provided to produce the appropriate input definition files and the jobscript files for strong scaling tests up to 4096 MPI processes. To produce the necessary files for Nmpi process job, run the following command on a system which supports Python.
./makeDef_large.py Nmpi
After this script is run, there will be a jobscript "job_mpi${Nmpi}" and a set of input definition files named "*.def". The number of groups are set as 64, so jobs with the multiple of 64 MPIs are good choices. A job with 64 MPIs will take about one hour on K computer.
The number of OpenMP threads can be set independent of the input difinition files. If the number of threads exceeds (NSPGaussLeg*NMPTrans)/NSplitSize, then the thread load imbalance may occur. By default, the job.sh file created by above makeDef_large.py sets the threads value as:
export OMP_NUM_THREADS=8
mVMC-mini is a subset of mVMC full application, and essentially retains the same feature as mVMC. The original mVMC has the following features. Using the multi-variable variational parameters, mVMC analyzes the physical characteristics of the strongly correlated electron systems, by configuring the variational wave functions closely representing the ground state of such electron systems. It executes the Monte Carlo sampling of the real space configuration of the electrons.
The variational wave function in mVMC is composed of three parts, the singlet paring wave function, correlation factors, and quantum-number projections. mVMC computes physical properties of the variational wave function in the strongly correlated electron systems, such as energy, magnetism and superconductivity. To carry out such computation, it computes Pfaffian of the skew-symmetric matrix to obtain the inner product between the paring wave function and the real space electron configuration.
In optimizing the variational parameters, it applies the stochastic reconfiguration method, which is more stable than the steepest descent method. The Monte Carlo sampling and the matrix computation process in the optimization process are both parallelized.
For detail explanation of mVMC, refer to the paper: Tahara D and Imada M, J. Phys. Soc. Jpn. 77, 114701 (2008) "Variational Monte Carlo method combined with quantum-number projection and multi-variable optimization."
Contact point of the original mVMC: Dr. Satoshi Morita morita@issp.u-tokyo.ac.jp
The milestone performance of the 10,000 atom system simulation using mVMC is expected to be 8 hours on exa-scale platform. Such milestone is under review.