StarPU MORSE Nightly Tests Archive

• dgemm, sgemm and spotrf miriel benchmarks are run on OPA miriel nodes.
• other spotrf benchmarks are run on sirocco nodes from 2017, december 15th (sequential benchmarks used to be run on mirage nodes).

Execution on queue routage with 4 node(s) (-C "haswell&sirocco") of the following script:

printenv | grep SLURM 
printenv | grep STARPU_ 
module list 
ulimit -c 0 
export STARPU_WORKER_STATS=1 
export STARPU_CALIBRATE=2 
export STARPU_LIMIT_CPU_MEM=120000 
export STARPU_LIMIT_MIN_SUBMITTED_TASKS=15000 
export STARPU_LIMIT_MAX_SUBMITTED_TASKS=16000 
export p=$(perl -e 'print int(sqrt('$SLURM_NNODES')), "n"') 
export nbgpus=4 
procs=$(echo $SLURM_JOB_CPUS_PER_NODE | awk -F'(' '{print $1}') 
export nbthreads=$(($procs - 2 -$nbgpus)) 
export niter=${NITER:-30} 
type mpiexec 
mpiexec -DPATH=${PATH} -DLD_LIBRARY_PATH=${LD_LIBRARY_PATH} -np $SLURM_NNODES hostname 
mpiexec -DPATH=${PATH} -DLD_LIBRARY_PATH=${LD_LIBRARY_PATH} -np $SLURM_NNODES $PWD/testing/chameleon_stesting \ 
   --op potrf --check --niter $niter --ib 96 --nb 1440 --n 144000:144000:14400 --P $p --threads $nbthreads \ 
   --gpus $nbgpus 

• results_master_HEAD_spotrf_mpi.dat
• results_master_HEAD_spotrf_mpi_nmad.dat
• results_master_HEAD_spotrf_mpi_mad_mpi.dat
• results_master_HEAD_spotrf_mpi_sim.dat