Compare Aurora and Licallo execution times
The aim is to run the same simulations on Aurora and Licallo to check both results and runtime
- Concerning results we are ok
- Concerning runtime we are a bit puzzled:
On Licallo we used:
#SBATCH --partition=fdr
#SBATCH --nodes=4
#SBATCH --exclusive
#SBATCH --ntasks-per-node=2
#SBATCH --cpus-per-task=1
and have 4 minutes for 5 outputs (/scratch/elena/TestMPIfdr)
on Aurora:
#SBATCH --nodes=4
#SBATCH --ntasks-per-node=2
#SBATCH --cpus-per-task=1
#SBATCH --exclusive. this reserves the full node,
something from aurora documentation
Tue Jul 9 18:27:46 CEST 2019
Tue Jul 9 18:58:31 CEST 2019
31 min
**Considering that fdr:**
Proc -> bi cpus Intel Xeon IvyBridge E5-2670 v2 à 2.50GHz ; 2 x 10 coeurs
Ram -> 64 Go
Infiniband-> FDR-56Gb/s
**and Aurora:**
CPU 2 Intel Xeon E5-2650 v3 (2.3 Ghz, 10-core)
Memory 64 GB (3.2 GB/core)
Local disk 1.7 TB, temporary directory given by $SNIC_TMP or $TMPDIR
Interconnect 4xFDR InfiniBand
may be InfiniBand is slower (I have no idea) but we didn't expect a factor 7 longer on Aurora,
@lainm do you have any idea?