Arkansas High Performace Computing Center [hpcwiki]

Version 5.1 Compilation

With Intel compiler and either OpenMPI or MVAPICH2:

OpenMPI:
DFLAGS         = -D__INTEL -D__FFTW3 -D__MPI -D__PARA -D__SCALAPACK $(MANUAL_DFLAGS)
IFLAGS         = -I../include
MPIF90         = mpif90
CFLAGS         = -O3 -xSSE2 -axavx $(DFLAGS) $(IFLAGS)
F90FLAGS       = $(FFLAGS) -nomodule -fpp $(FDFLAGS) $(IFLAGS) $(MODFLAGS)
FFLAGS         = -O2 -xSSE2 -axavx -assume byterecl -g -traceback -par-report0 -vec-report0
FFLAGS_NOOPT   = -O0 -assume byterecl -g -traceback
FFLAGS_NOMAIN  = -nofor_main
LD             = mpif90
LDFLAGS        = -static-intel 
SCALAPACK_LIBS = -lmkl_scalapack_lp64 -lmkl_blacs_openmpi_lp64
FFT_LIBS       = -L ${MKL_ROOT}/interfaces/fftw3xf -lfftw3xf_intel

MVAPICH2: same except
SCALAPACK_LIBS = -lmkl_scalapack_lp64 -lmkl_blacs_intelmpi_lp64

trestles: same except
no -axavx (though an "optional" code path, it makes the program fail on AMD)

Benchmarks

We run AUSURF112 from Espresso Benchmarks and compare with Glenn Lockwood who ran the AUSURF112 benchmark on SDSC Comet and on the Trestles system when it was at SDSC. Unfortunately, the AUSURF112 benchmark generally ends the simulation with convergence NOT achieved after 2 iterations: stopping, but it does so fairly repeatably so may be timed.

OpenMPI:
module load intel/14.0.3 mkl/14.0.3 openmpi/1.8.8 
mpirun -np 64  -machinefile $PBS_NODEFILE -x LD_LIBRARY_PATH \
/share/apps/espresso/espresso-5.1-intel-openmpi/bin/pw.x -npools 1 <ausurf.in
MVAPICH2:
module load intel/14.0.3 mkl/14.0.3 mvapich2/2.1
mpirun -np 64  -machinefile $PBS_NODEFILE \
/share/apps/espresso/espresso-5.1-intel-mvapich2/bin/pw.x -npools 1 <ausurf.in

The tables shows Lockwood's and our times. We add 32×4 core runs for Trestles as we think node-to-node is the more representative comparison. Our newer versions of OpenMPI show better results on the almost-identical hardware than Lockwood.

(1) Fails with error charge is wrong.

Notes

Each run fails with error messages (depending on MPI type) and RC 1 after terminating normally according to the log. This appears harmless:

   This run was terminated on:  13: 2:44  11Nov2015            
=------------------------------------------------------------------------------=
   JOB DONE.
=------------------------------------------------------------------------------=
-------------------------------------------------------
Primary job  terminated normally, but 1 process returned
a non-zero exit code.. Per user-direction, the job has been aborted.
-------------------------------------------------------
------------------------------------------------------------
A process or daemon was unable to complete a TCP connection
to another process:
etc.

Continuing Work

ELPA in newer versions of Espresso is reportedly faster than Scalapack.

OpenMPI threading.

MKL threading.

FFTW fft vs. Intel fft on AMD.

On Trestles with Intel tools. It's difficult to find a combination of versions that is new enough to compile qe-6.6 (compiler > 18) yet will still produce a binary that runs on AMD Bulldozer (mkl <20) while avoiding most mkl bugs (mkl > 18).

module load intel/18.0.2 mkl/19.0.5 impi/17.0.4
MKL_NUM_THREADS=# OMP_NUM_THREADS=# mpirun -np ## -machinefile machinefile /share/apps/espresso/espresso-6.6-intel-impi-mkl-trestles/bin/pw.x <ausurf.in

q-e appears to be a code that does not like mpi threads x OMP threads > physical cores. Performance on two trestles nodes is better than the previous 5.1 benchmarks, but it doesn't scale past two nodes on this small problem.

Cores Node  type  #mpi  #nodes   #OMP #MKL  WALL
32  Trestles AMD    32       1      1    1  12m11s
32  Trestles AMD    32       1      2    1  50m41s
32  Trestles AMD    32       1      1    2  >14m
32  Trestles AMD    16       1      2    1  >14m
32  Trestles AMD    64       2      1    1   8m33s
32  Trestles AMD   128       4      1    1   8m42s
32  6130   Intel    32       1      1    1   2m35s
32  6130   Intel    32       1      1    1   2m33s *
32  6130   Intel    32       1      1    1   2m58s ***
48  7402     AMD    48       1      1    1   1m58s
48  7402     AMD    24       1      2    2   3m32s
48  7402     AMD    96       1      2    2   2m20s  **

• using a better optimized version for more modern machines, which doesn't seem to help much.

tested for number of hardware threads, which is negative for performance vs. number of physical cores * 6.5 version.

Install script

OMP="--enable-openmp"
VERSION=6.6
./install/configure MPIF90=mpiifort F90=ifort F77=ifort FC=ifort CC=icc \
SCALAPACK_LIBS="-L$MKLROOT/lib/intel64 -lmkl_scalapack_lp64 -lmkl_blacs_intelmpi_lp64" \
LAPACK_LIBS="-L$MKLROOT/lib/intel64 -lmkl_lapack95_lp64 -lmkl_blas95_lp64" \
BLAS_LIBS="-lmkl_intel_lp64  -lmkl_intel_thread -lmkl_core -liomp5 -thread" \
FFT_LIBS="-L$MKLROOT/interfaces/fftw3xf -lfftw3xf_intel" \
FFLAGS="-O3 -xHOST -D__INTEL -D__GNUC__ -D__FFTW3 -D__MPI -D__PARA -D__SCALAPACK \
-assume byterecl -I$MKLROOT/include/fftw" \
CFLAGS="-O3 -xHOST -D__INTEL -D__GNUC__ -D__FFTW3 -D__MPI -D__PARA -D__SCALAPACK" \
--with-hdf5=/share/apps/hdf5/1.10.5/intel/impi -with-scalapack=intel \
--enable-parallel $OMP --prefix=/share/apps/espresso/espresso-$VERSION-intel-impi-mkl-trestles

QE 6.8 are 7.1 installed with two versions compiled with the Intel compiler (“skylake” for Intel and “bulldozer” for AMD). “skylake” uses -xHOST and is compiled on Pinnacle I. “bulldozer” uses -msse3 -axsse3,sse4.2,AVX,core-AVX2,CORE-AVX512 and is compiled on Trestles, so should work at some speed on all systems. Both use module load intel/19.0.5 mkl/19.0.5 impi/17.0.4. The impi/19.0.5 module causes a fault on the AMD platforms for unknown reasons. The “skylake” binary causes a fault on the AMD platforms because of a single AVX512 code path. export MKL_DEBUG_CPU_TYPE=5 as set on AMD by module mkl<20 causes a fault on the AMD platforms (failure on Trestles, wrong answer on Pinnacle II). For AMD, explicitly set after the module load:

export MKL_DEBUG_CPU_TYPE=0

A small pw.x input was used to allow some parameter sweeps.

Single Node Results

{OMP_NUM_THREADS=1|2} time mpirun -np {16|32|64} \
/share/apps/espresso/espresso-{6.8|7.1}-intel-impi-mkl-{skylake|bulldozer}/bin/pw.x \
-nk {1|4|8|16} <scf.in >log

System   Cores        CPU  QE version compile #mpi #OMP #MKL  -nk Wall Time(s) 

Pinnacle I  32 Intel 6130      6.8    skylake   32    1    1    1   20.5
Pinnacle I  32 Intel 6130      6.8    skylake   32    1    1    4   13.6
Pinnacle I  32 Intel 6130      6.8    skylake   32    1    1    8    9.5
Pinnacle I  32 Intel 6130      6.8    skylake   32    1    1   16   14.3

Pinnacle I  32 Intel 6130      6.8    skylake   16    2    1    4   >360
Pinnacle I  32 Intel 6130      6.8  bulldozer   32    1    1    8   22.0 

Trestles    32   AMD 6136      6.8  bulldozer   32    1    1    8   17.5 
Trestles    32   AMD 6136      6.8  bulldozer   32    1    1    4   16.2 
Trestles    32   AMD 6136      7.1  bulldozer   32    1    1    4   15.3 

Pinnacle II 64   AMD 7543      6.8  bulldozer   64    1    1    4    4.9 
Pinnacle II 64   AMD 7543      6.8  bulldozer   64    1    1    8    4.0 
Pinnacle II 64   AMD 7543      6.8  bulldozer   64    1    1   16    4.5 
Pinnacle II 64   AMD 7543      6.8  bulldozer   32    1    1    8    6.0 
Pinnacle II 64   AMD 7543      7.1  bulldozer   64    1    1    8    4.2 

Conclusions for this sample program:

-nk 8 is best on Pinnacle I & II platforms, -nk 4 on Trestles.

QE 7.1 is slightly slower than 6.8 on Pinnacle I & II and slightly faster on Trestles.

OMP_NUM_THREADS>1 is a very substantial slowdown versus 1 or unset.

The bulldozer version runs on Intel but is significantly slower than the skylake version (other older platforms such as E5 condo nodes won't run AVX512 codes and would probably benefit from their own version).

The usually zero-wait Trestles system has relatively good performance on QE if shared memory (64 GB) allows (relatively meaning ~1/2 of Pinnacle I performance when for some programs it is ~1/5)