From bb5c16b97f1c40b7c131eae08ef58f989e2c07a4 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Thu, 3 Mar 2022 17:00:37 -0500
Subject: [PATCH] New scripts

---
 systems/Crusher/dwf.slurm     | 20 ++++++++++----------
 systems/Crusher/dwf8.slurm    | 33 +++++++++++++++++++++++++++------
 systems/Crusher/mpiwrapper.sh |  5 +++--
 3 files changed, 40 insertions(+), 18 deletions(-)

diff --git a/systems/Crusher/dwf.slurm b/systems/Crusher/dwf.slurm
index 286615ef..8742ed76 100644
--- a/systems/Crusher/dwf.slurm
+++ b/systems/Crusher/dwf.slurm
@@ -3,28 +3,28 @@
 #SBATCH -A LGT104
 #SBATCH -t 01:00:00
 ##SBATCH -U openmpThu
-##SBATCH -p ecp
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
 #SBATCH -e DWF.%J
 #SBATCH -N 1
-#SBATCH -n 1
-#SBATCH --exclusive  
+#SBATCH -n 8
+#SBATCH --exclusive
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
 
 DIR=.
 module list
-#export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-#export MPICH_SMP_SINGLE_COPY_MODE=NONE
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
 #export MPICH_SMP_SINGLE_COPY_MODE=CMA
 export OMP_NUM_THREADS=1
 
-AT=8
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
 
-PARAMS=" --accelerator-threads ${AT} --grid 24.24.24.24 --shm-mpi 0 --mpi 1.1.1.1"
-
-srun --gpus-per-task 1 -n1 ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+PARAMS=" --accelerator-threads 16 --grid 32.32.32.256 --mpi 1.1.1.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+srun --gpus-per-task 1 -n8  ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+
 
 
diff --git a/systems/Crusher/dwf8.slurm b/systems/Crusher/dwf8.slurm
index 30e83fff..64572142 100644
--- a/systems/Crusher/dwf8.slurm
+++ b/systems/Crusher/dwf8.slurm
@@ -6,22 +6,43 @@
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
 #SBATCH -e DWF.%J
-#SBATCH -N 1
-#SBATCH -n 8
+#SBATCH -N 8
+#SBATCH -n 64
 #SBATCH --exclusive
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
 
 DIR=.
 module list
+export MPICH_OFI_NIC_POLICY=GPU
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-#export MPICH_SMP_SINGLE_COPY_MODE=NONE
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
 #export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
 export OMP_NUM_THREADS=1
 
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
-PARAMS=" --accelerator-threads 8 --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
 
-srun --gpus-per-task 1 -n8 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.256 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.256.8node
+
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 1"
+echo $PARAMS
+srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node.shm0
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 1"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node_shm0
+
 
 
diff --git a/systems/Crusher/mpiwrapper.sh b/systems/Crusher/mpiwrapper.sh
index 76c4e364..f6a56698 100755
--- a/systems/Crusher/mpiwrapper.sh
+++ b/systems/Crusher/mpiwrapper.sh
@@ -1,10 +1,11 @@
 #!/bin/bash
 
 lrank=$SLURM_LOCALID
+lgpu=(0 1 2 3 7 6 5 4)
 
-export ROCR_VISIBLE_DEVICES=$SLURM_LOCALID
+export ROCR_VISIBLE_DEVICES=${lgpu[$lrank]}
 
-echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES binding=$BINDING"
+echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES "
 
 $*