StarPU

StarPU Morse Nightly Tests

Buildbot

Buildbot

Contents

Tests for morse/morse_seq.txt

Tests for trunk/chameleon #HEAD (Profile seq)

Execution on queue court_mirage with 1 nodes (--gres=gpu:3 -C M2070 -n 12) of the following script
	export STARPU_WORKER_STATS=1
	export STARPU_CALIBRATE=2
        ./timing/time_spotrf_tile --gpus=3 --threads=9 --niter=30 --ib=96 --nb=960 --n_range=48000:48000:9600 --check
date morse_branch morse_scm starpu_svn output file error_file M N K/NRHS seconds Gflop/s Deviation
2017-10-11 trunk/chameleon f200958 22317 output error 48000 48000 1 19.688 1872.50 9.89
2017-10-10 trunk/chameleon f200958 22307 output error 48000 48000 1 19.662 1875.01 11.31
2017-10-07 trunk/chameleon f200958 22302 output error 48000 48000 1 19.700 1871.42 13.63
2017-10-06 trunk/chameleon f200958 22288 output error 48000 48000 1 19.694 1871.94 9.86
2017-10-05 trunk/chameleon f200958 22281 output error 48000 48000 1 19.552 1885.49 5.27
2017-10-04 trunk/chameleon f200958 22270 output error 48000 48000 1 19.649 1876.20 9.85
2017-09-20 trunk/chameleon 6c90018 22161 output error 48000 48000 1 19.666 1874.63 9.48
2017-09-19 trunk/chameleon 6c90018 22157 output error 48000 48000 1 19.660 1875.21 10.24
2017-09-16 trunk/chameleon 6c90018 22154 output error 48000 48000 1 19.668 1874.45 8.00
2017-09-15 trunk/chameleon 6c90018 22153 output error 48000 48000 1 19.679 1873.41 9.84
See also the tests archive.

Tests for trunk/chameleon #HEAD (Profile seq_sim)

Execution on queue court with 1 nodes () of the following script
	export STARPU_HOME="$prefix_dir/../morse/trunk/chameleon/simucore/perfmodels/"
	export STARPU_HOSTNAME=mirage
	export STARPU_CALIBRATE=0
	export STARPU_WORKER_STATS=1
        time ./timing/time_spotrf_tile --nowarmup --gpus=3 --threads=9 --niter=30 --nb=960 --ib=96 --n_range=48000:48000:9600
date morse_branch morse_scm starpu_svn output file error_file M N K/NRHS seconds Gflop/s Deviation
2017-10-11 trunk/chameleon f200958 22317 output error 48000 48000 1 19.894 1853.13 1.93
2017-10-10 trunk/chameleon f200958 22307 output error 48000 48000 1 19.892 1853.23 3.28
2017-10-07 trunk/chameleon f200958 22302 output error 48000 48000 1 19.897 1852.82 2.99
2017-10-06 trunk/chameleon f200958 22288 output error 48000 48000 1 19.894 1853.12 2.34
2017-10-05 trunk/chameleon f200958 22281 output error 48000 48000 1 19.889 1853.53 2.38
2017-10-04 trunk/chameleon f200958 22270 output error 48000 48000 1 19.892 1853.25 2.78
2017-10-03 trunk/chameleon e4c0d24 22264 output error 48000 48000 1 19.890 1853.44 3.21
2017-09-30 trunk/chameleon deef871 22251 output error 48000 48000 1 19.895 1853.00 3.00
2017-09-29 trunk/chameleon e3daf9a 22240 output error 48000 48000 1 19.899 1852.65 2.47
2017-09-28 trunk/chameleon e3daf9a 22234 output error 48000 48000 1 19.897 1852.76 2.94
See also the tests archive.

Tests for morse/morse_mpi.txt

Tests for trunk/chameleon #HEAD (Profile mpi)

Execution on queue court_sirocco with 4 nodes (-C Haswell --gres=gpu:4) of the following script
        printenv | grep SLURM
        printenv | grep STARPU_
        module list

	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
	# Compute P to set PxQ. Use square grid (Chameleon is unsafe with a non square grid), so P = Q = sqrt(number of nodes).
	export p=$(perl -e 'print int(sqrt('$SLURM_JOB_NUM_NODES')), "\n"')
	# Number of threads = number of processors per node - 1 - number of GPUS. We need to make sure the CPU managing GPU devices are not counted.
        export nbgpus=4
	export nbthreads=$(($SLURM_CPUS_ON_NODE - 1 - $nbgpus))
	# Start execution by giving the number of threads and the grid size. The problem size is set according to the given number of nodes.
	mpiexec -np $SLURM_JOB_NUM_NODES -pernode ./timing/time_spotrf_tile --niter=30 --ib=96 --nb=1440 --n_range=144000:144000:14400 --P=$p --threads=$nbthreads --gpus=$nbgpus
date morse_branch morse_scm starpu_svn output file error_file M N K/NRHS seconds Gflop/s Deviation
2017-10-12 trunk/chameleon f200958 22317 output error 144000 144000 1 31.073 32032.88 59.67
2017-10-07 trunk/chameleon f200958 22302 output error 144000 144000 1 31.259 31841.57 72.59
2017-10-06 trunk/chameleon f200958 22288 output error 144000 144000 1 31.021 32086.50 49.95
2017-10-05 trunk/chameleon f200958 22281 output error 144000 144000 1 31.164 31939.40 75.85
2017-10-04 trunk/chameleon f200958 22270 output error 144000 144000 1 31.045 32061.45 43.75
2017-09-19 trunk/chameleon 6c90018 22157 output error 144000 144000 1 34.517 28870.62 988.47
2017-01-18 trunk/chameleon 4b3a2f8 19639 output error 144000 144000 1 31.476 31622.60 168.06
2016-10-27 trunk/chameleon 3167 19171 output error 144000 144000 1 35.292 28205.25 262.43
2016-10-26 trunk/chameleon 3167 19163 output error 144000 144000 1 35.401 28118.16 216.52
2016-10-25 trunk/chameleon 3167 19152 output error 144000 144000 1 36.965 26951.23 763.28
See also the tests archive.

Tests for trunk/chameleon #HEAD (Profile mpi_sim)

Execution on queue court with 1 nodes () of the following script
        printenv | grep SLURM
        printenv | grep STARPU_
        module list

	export STARPU_HOME="$prefix_dir/../morse/trunk/chameleon/simucore/perfmodels/"
	export STARPU_HOSTNAME=sirocco
	export STARPU_CALIBRATE=0
	export STARPU_WORKER_STATS=1
        export STARPU_LIMIT_CPU_MEM=120000
        export STARPU_LIMIT_MIN_SUBMITTED_TASKS=15000
        export STARPU_LIMIT_MAX_SUBMITTED_TASKS=16000
        export STARPU_MALLOC_SIMULATION_FOLD=2
	# Hardcoder le parametrage de la soumission MPI
	SLURM_JOB_NUM_NODES=4
	SLURM_CPUS_ON_NODE=24
	# Calcul de P pour déterminer PxQ. On prend une grille carrée (Chameleon bug plus souvent avec une grille non carrée), donc P = Q = sqrt(nombre de nœuds).
	p=$(perl -e 'print int(sqrt('$SLURM_JOB_NUM_NODES')), "\n"')
	# Le nombre de threads est le nombre de processeurs par nœud moins 1, moins le nombre de GPUs. Vu qu'on donne cette valeur à la main, il faut retrancher les threads pour les workers GPU soi-même.
        nbgpus=4
	nbthreads=$(($SLURM_CPUS_ON_NODE - 1 - $nbgpus))
	# Ligne de compilation : on ajoute les paramètres threads et p au lancement du test. La taille du problème sera à configurer en fonction du nombre de nœuds choisi.
	time $prefix_dir/bin/starpu_smpirun -np $SLURM_JOB_NUM_NODES -platform $HOME/.starpu/cluster.xml -hostfile $HOME/.starpu/hostfile --cfg=surf/precision:0.000010 --cfg=maxmin/precision:0.000010 --cfg=network/model:IB --cfg=smpi/test:0.000010 ./timing/time_spotrf_tile --nowarmup --ib=96 --nb=1440 --n_range=144000:144000:14400 --P=$p --gpus=$nbgpus
date morse_branch morse_scm starpu_svn output file error_file M N K/NRHS seconds Gflop/s Deviation
2017-10-12 trunk/chameleon f200958 22317 output error 144000 144000 1 30.365 32778.87 0.00
2017-10-10 trunk/chameleon f200958 22307 output error 144000 144000 1 30.365 32778.87 0.00
2017-10-07 trunk/chameleon f200958 22302 output error 144000 144000 1 30.339 32807.05 0.00
2017-10-06 trunk/chameleon f200958 22288 output error 144000 144000 1 30.316 32832.09 0.00
2017-10-05 trunk/chameleon f200958 22281 output error 144000 144000 1 30.339 32807.05 0.00
2017-10-04 trunk/chameleon f200958 22270 output error 144000 144000 1 30.339 32807.05 0.00
2017-10-03 trunk/chameleon e4c0d24 22264 output error 144000 144000 1 30.339 32807.05 0.00
2017-09-30 trunk/chameleon deef871 22251 output error 144000 144000 1 30.339 32807.05 0.00
2017-09-29 trunk/chameleon e3daf9a 22240 output error 144000 144000 1 30.339 32807.05 0.00
2017-09-28 trunk/chameleon e3daf9a 22234 output error 144000 144000 1 30.339 32807.05 0.00
See also the tests archive.

Last updated on 2017/10/11 at 04:05.