Differenze

Queste sono le differenze tra la revisione selezionata e la versione attuale della pagina.

--- calcoloscientifico:benchmarks [04/06/2018 12:28] – [Job MPI with IntelMPI] roberto.alfieri
+++ calcoloscientifico:benchmarks [17/04/2020 15:44] (versione attuale) – roberto.alfieri
@@ Linea 1: / Linea 1: @@
-<html><center><b> <font face="Cursive" color="#606060" size="2">
-Progetto HPC <em>- Calcolo Scientifico dell'Università e dell'INFN di Parma -</em>
-<br><font size="5"> Benchmarks </font></font> </b></center></html>
+<html><center><b>
+<font face="Arial" color="#606060" size="2">
+ Scientific Computing of the  University of Parma, <em> in collaboration with INFN </em>
+<br><font size="5"> HPC Benchmark </font></font> </b></center><p></html>
@@ Linea 27: / Linea 31: @@
 <code bash>
 #!/bin/sh
-# Two processes on the same node BDW.
+# Due processi su stesso host stesso socket
 #SBATCH --output=%x.o%j
@@ Linea 39: / Linea 42: @@
 #SBATCH  --account=T_HPC18A
+##SBATCH  --reservation corso_hpc18a_7
 ## Print the list of the assigned resources
 echo "#SLURM_JOB_NODELIST: $SLURM_JOB_NODELIST"
-module load  gnu openmpi imb
+module load  intel intelmpi
-CMD="mpirun   IMB-MPI1  pingpong  -off_cache -1"
+CMD="mpirun   /hpc/group/T_HPC18A/bin/IMB-MPI1  pingpong  -off_cache -1"
 echo "# $CMD"
-eval $CMD > IMB-N1.dat
+eval $CMD > IMB-N1-BDW.dat
 </code>
@@ Linea 53: / Linea 57: @@
 <code bash>
 #!/bin/sh
 # Due processi su stesso host stesso socket
@@ Linea 70: / Linea 73: @@
 echo "#SLURM_JOB_NODELIST: $SLURM_JOB_NODELIST"
-module load  gnu openmpi imb
+module load  intel intelmpi
-mpirun hostname
+CMD="mpirun   /hpc/group/T_HPC18A/bin/IMB-MPI1.knl  pingpong  -off_cache -1"
-CMD="mpirun   IMB-MPI1  pingpong  -off_cache -1"
 echo "# $CMD"
 eval $CMD > IMB-N1-KNL.dat
@@ Linea 100: / Linea 101: @@
 echo "#SLURM_JOB_NODELIST: $SLURM_JOB_NODELIST"
-module load  gnu openmpi imb
+module load  intel intelmpi
-#mpirun hostname
+mpirun hostname
-CMD="mpirun   IMB-MPI1  pingpong  -off_cache -1"
+CMD="mpirun   /hpc/group/T_HPC18A/bin/IMB-MPI1  pingpong  -off_cache -1"
 echo "# $CMD"
-eval $CMD > IMB-N2.dat
+eval $CMD > IMB-N2-opa.dat
+export  I_MPI_FABRICS=shm:tcp
+CMD="mpirun   /hpc/group/T_HPC18A/bin/IMB-MPI1  pingpong  -off_cache -1"
+echo "# $CMD"
+eval $CMD > IMB-N2-tcp.dat
 </code>
-Results 2018 with GNU  compiler:
+Results 2018 (Intel compiler):
 {{:calcoloscientifico:imb.png?200|}}
-Results 2017 with Intel compiler:
+Results 2017  (Intel compiler):
 {{:calcoloscientifico:imb-mpi1-n1.png?200|}}
@@ Linea 122: / Linea 131: @@
 ===== mpi_latency, mpi_bandwidth =====
-==== Job MPI with IntelMPI ====
   module load intel intelmpi
+  #module load gnu openmpi
   cp /hpc/share/samples/mpi/mpi_latency.c   .
   cp /hpc/share/samples/mpi/mpi_bandwidth.c .
@@ Linea 132: / Linea 141: @@
-Script ''mpi_lat_band.sh '':
+Script ''mpi_lat_band.slurm '':
 <code bash>
@@ Linea 140: / Linea 149: @@
 #SBATCH --partition=bdw
-#SBATCH  -N2 --tasks-per-node=1
+### 2 nodi (OPA o TCP)
+#SBATCH  -N2 --tasks-per-node=1
+### 1 nodo (SHM)
+##SBATCH  -N1 --tasks-per-node=2
 #SBATCH  --exclusive
+#SBATCH  --account=T_2018_HPCCALCPAR
+##SBATCH  --reservation hpcprogpar_20190517
 ## Print the list of the assigned resources
@@ Linea 152: / Linea 168: @@
 export  I_MPI_FABRICS=shm:tcp
 mpirun  mpi_latency    > mpi_latency_TCP.dat
 mpirun  mpi_bandwidth  > mpi_bandwidth_TCP.dat
@@ Linea 159: / Linea 173: @@
 </code>
-Results 2018:
+Results 2019 (INTEL compiler):
-^  Latency  ^^  Bandwidth  ^^
+^  Latency (micros.)  ^^^  Bandwidth (MB/s)  ^^^
-^  OPA   ^  TCP  ^  OPA  ^  TCP  ^
+^  SHM  ^  OPA  ^  TCP  ^  SHM  ^  OPA  ^  TCP  ^
-|  2 us  |  86 us ^ 6600 MB/s |  117 MB/s  |
+|  3  |  3  |  86  |  7170  |  6600  |  117   |
 ===== NBODY =====
@@ Linea 361: / Linea 374: @@
-==== Run GPU ====
-TODO