Kernel-benchmark

Benchmark of kernels in ChASE and others

Module loaded

module load GCC
module load CMake/3.14.0
module load ParaStationMPI/5.2.2-1
module load Boost/1.69.0-Python-2.7.16
module load imkl 

Compile

mkdir build
cd build
cmake ..
make -j 

HEEVD function

Execution

multi-threading version

./HEEVD/heevd.exe --s=${SIZE} --omp:threads=${NUM_THREAD}

where ${SIZE} specifies the matrix size, and by ${NUM_THREAD} parse the number of threads to be used.

MPI version

MPI version runs independantly mutli-threading HEEVD on each proc.

mpirun -np ${PROCS} ./HEEVD/heevd_mpi.exe --s=${SIZE} --omp:threads=${NUM_THREAD}

Improved MPI version

In this version, on each node, HEEVD is executed on only one proc in the node, and then the result is broadcasted to other procs within the same node

mpirun -np ${PROCS} ./HEEVD/heevd_shm_mpi.exe --s=${SIZE} --omp:threads=${NUM_THREAD}

Update QR

Paramter

This is a simulator of the QR inside ChASE. For the simulation, several parameters are determined:

• ${M}: Row number of matrix to be factored, which should be equal to the matrix size in ChASE

• ${NEV}: Number of eigenvector to be computed

• ${NEX}: Extra number of vectors to be used to comute eigenpairs

• ${LOCKED}: Number of eigenparis locked in each step of iteration. This parameter should be dynamic in ChASE, but here I set it to be static for simplication.

• ${THREADS}: the number of threads to be used

Execution

Chase-like update QR

./updateQR/chaseQR.exe --m=$M --nev=${NEV} --nex=${NEX} --locked=${LOCKED} --omp:threads=${THREADS}

New proposed update QR

./updateQR/updateQR.exe --m=$M --nev=${NEV} --nex=${NEX} --locked=${LOCKED} --omp:threads=${THREADS}