run-elpa

#!/bin/bash

# cori gpu
# will need to issue the following sometime before running benchmark_elpa
#
# module purge ; module load esslurm; module load intel/19.0.3.199; module load cuda/10.2.89; module load openmpi/4.0.3

# request one compute node on cori gpu
# salloc -C gpu -N 1 -G 8 -t 30 -A m1759 -q special --exclusive

#

export OMP_NUM_THREADS=1

# ELPA requires block sizes of powers of 2
# example to run on all 8 GPUs on a compute node

srun -N 1 -n 8 -c 10 --cpu_bind=cores ./benchmark_elpa 4096 4 2 32






# ignore warnings such as the following:
#
#[1597242445.241043] [cgpu04:25499:0]    ucp_context.c:586  UCX  WARN  transport 'cuda_copy' is not available
#[1597242445.241050] [cgpu04:25499:0]    ucp_context.c:586  UCX  WARN  transport 'gdr_copy' is not available
#[1597242445.241051] [cgpu04:25499:0]    ucp_context.c:586  UCX  WARN  transport 'cuda_ipc' is not available