This project contains code speed comparisons between different python and fortran programs run
on the compute nodes of the spinon cluster of the F1 department at IJS. I am particularly interested
in comparing performance of intelpython2/2018-u1 and Anaconda2/5.3.0 in diagonalization of symmetric
(non-sparse, at the moment) matrices, as well as the performance of the syev diagonalization subroutine
from intel's MKL library.
Additionally, comparison of the naive python for loop implementation and the speeded-up numba version
will also be available in the future.
#!/bin/bash
name_source=$1
compiler=ifort
name_object=${name_source/.f90/.o} #substitution
echo "Compiling source code:" $name_source
libs1="-O3 -mcmodel=medium -fpic -heap-arrays 0 -align all -funroll-loops -qopenmp -mkl=parallel -lmkl_core -lmkl_lapack95_lp64 -I${MKLROOT}/include/intel64/lp64 -I${MKLROOT}/include"
$compiler $name_source $libs1 -o $name_object
The code below submits jobs to the spinon cluster. Diagonalization of different symmetric random matrices is then performed, with the largest ones having size \(N=2^14=16384\).
#!/bin/bash
#SBATCH --time=2:59:00
#SBATCH --mem=4000
#SBATCH -c 8
#SBATCH --exclude=compute-1-[1,2,5-16,18-22,24,26,27,28]
#SBATCH --output=./Logs/log%J.out
#SBATCH --error=./Logs/log%J.e
echo ${SLURM_CPUS_PER_TASK}
source deactivate
module purge
module load intelpython2/2018-u1
export MKL_NUM_THREADS=${SLURM_CPUS_PER_TASK}
which python2.7
srun python2.7 python_diagonalization.py 0 14
module purge
module load Anaconda2/5.3.0
export MKL_NUM_THREADS=${SLURM_CPUS_PER_TASK}
which python2.7
srun python2.7 python_diagonalization.py 0 14
module purge
export MKL_NUM_THREADS=${SLURM_CPUS_PER_TASK}
module load intel/2018b
srun ./fortran_diagonalization.o 14
SLURM logs are stored in ./Logs, diagonalization and loop benchmarks are stored in ./diag_benchmarkand loop_benchmark folders, respectively. intel and anaconda filename prefixes stand for different python modules used.