HPC/Applications/lammps

Benchmark

Using a sample workload from Sanket ("run9"), I tested various OpenMPI options on both node types.

LAMMPS performs best on gen2 nodes without extra options, and pretty well on gen1 nodes over ethernet(!).

Job tag	Node type	Interconnect	Additional OpenMPI options	Relative speed (1000 steps/3 hours)	Notes
gen1	gen1	IB	(none)	36
gen1srqpin	gen1	IB	-mca btl_openib_use_srq 1 -mca mpi_paffinity_alone 1	39
gen1eth	gen1	Ethernet	-mca btl self,tcp	44	fastest for gen1
gen2eth	gen2	Ethernet	-mca btl self,tcp	49
gen2srq	gen2	IB	-mca btl_openib_use_srq 1	59
gen2	gen2	IB	(none)	59	fastest for gen2

Sample job file gen1

#!/bin/bash
#PBS -l nodes=10:ppn=8:gen1
#PBS -l walltime=1:00:00:00
#PBS -N <jobname>
#PBS -A <account>
#
#PBS -o job.out
#PBS -e job.err
#PBS -m ea

# change into the directory where qsub will be executed
cd $PBS_O_WORKDIR

mpirun  -machinefile  $PBS_NODEFILE \
        -np $(wc -l < $PBS_NODEFILE) \
        -mca btl self,tcp \
        lmp_openmpi < lammps.in > lammps.out 2> lammps.err

Sample job file gen2

#!/bin/bash
#PBS -l nodes=10:ppn=8:gen2
#PBS -l walltime=1:00:00:00
#PBS -N <jobname>
#PBS -A <account>
#
#PBS -o job.out
#PBS -e job.err
#PBS -m ea

# change into the directory where qsub will be executed
cd $PBS_O_WORKDIR

mpirun  -machinefile  $PBS_NODEFILE \
        -np $(wc -l < $PBS_NODEFILE) \
        lmp_openmpi < lammps.in > lammps.out 2> lammps.err

MPI/OpenMP hybrid parallel runs

LAMMPS modules since 2012 are compiled with yes-user-omp, permitting multi-threaded runs of selected pair styles, and in particular MPI/OpenMP hybrid parallel runs.

Be careful how to allocate CPU cores on compute nodes. Note the following:

• The number of MPI tasks (call it ppn_mpi) running on a node is determined by the qsub ppn=... parameter.
• The number of threads that each MPI task runs with is determined by the environment variable OMP_NUM_THREADS, which is 1 by default on Carbon.
• The number of physical cores per node for gen1 and gen2 nodes is 8.
• gen2 nodes have hyperthreading active, meaning there are 16 logical cores per node. However:
  • My (stern) own benchmarks for a memory-intensive DFT program were underwhelming.
  • The LAMMPS OpenMP author reports the same (near the end of the section):

Using threads on hyper-threading enabled cores is usually counterproductive, as the cost in additional memory bandwidth requirements is not offset by the gain in CPU utilization through hyper-threading.

• • A mild benefit may be conferred for partial hyperthreading. Choose ppn_active such that

ppn_physical = 8 ≤ ppn_active ≤ ppn_logical = 16

• Reserve entire nodes by adding near the top of the job file:

#PBS -l naccesspolicy=SINGLEJOB

Sample job script for hybrid parallel runs

In summary, the job script's essential parts are:

#!/bin/bash
#PBS -l nodes=2:ppn=4
#PBS -l walltime=1:00:00
#PBS -l naccesspolicy=SINGLEJOB
...

ppn_mpi=$( uniq -c $PBS_NODEFILE | awk '{print $1; exit}' )	# grab first (and usually only) ppn value of the job
#ppn_active=8							# number of cores on first execution node
ppn_active=12							# experimental: 3-to-2 oversubscription of physical cores

OMP_NUM_THREADS=$(( ppn_active / ppn_mpi ))			# calculate number of threads available per MPI process (integer arithmetic!)

mpirun -x OMP_NUM_THREADS \
    -machinefile  $PBS_NODEFILE \
    -np $(wc -l < $PBS_NODEFILE) \
    lmp_openmpi \
	-sf omp \
	-in in.script

Diagnostic for hybrid parallel runs

• LAMMPS echoes it parallelization scheme first thing in the output:

LAMMPS (10 Feb 2012)
  using 3 OpenMP thread(s) per MPI task
...
  2 by 2 by 2 MPI processor grid
Lattice spacing in x,y,z = 3.52 4.97803 4.97803
...

and near the end:

Loop time of 11.473 on 24 procs (8 MPI x 3 OpenMP) for 100 steps with 32000 atoms

• To see if OpenMP is really active, log into a compute node while a job is running and run top. The %CPU field should be about OMP_NUM_THREADS × 100%

 PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                                                                             
8047 stern     25   0 4017m  33m 7540 R 301.8  0.1   1:41.60 lmp_openmpi                                                                                         
8044 stern     25   0 4017m  33m 7540 R 299.9  0.1   1:43.50 lmp_openmpi                                                                                         
8045 stern     25   0 4016m  33m 7536 R 299.9  0.1   1:43.47 lmp_openmpi                                                                                         
8046 stern     25   0 4017m  33m 7540 R 297.9  0.1   1:41.60 lmp_openmpi                                                                                         
4822 root      34  19     0    0    0 S  2.0  0.0 115:34.98 kipmi0                                                                                               

Bug

It appears at the moment that only the job's master node (first in $PBS_NODEFILE) runs multi-threaded – other nodes only get 100% CPU per MPI process, despite OMP_NUM_THREADS being correctly exported.

psuser
strings -a /proc/pid/environ | grep OMP_NUM_THREADS

OMP_NUM_THREADS=3

To learn more:

• HPC/Submitting_Jobs/Advanced node selection#Multithreading (OpenMP)
• Hints on using the OMP package
• Command-line options (explanation for -sf style or -suffix style)