10 May 2013

411. Attempt at OPENMP enabled NWChem 6.1.1 -- not successful...

Update 4 June 2013:
I might return to this later and have a look at how to make the parallel executable in the bin/LINUX64 folder.

Original post:
This is another addition to my growing list over unsuccessful, abandoned or only partially successful builds.
(see e.g.

In other words -- yes, it builds. But no, it is unusable.

I can build nwchem with openmp support, and it does run in parallel -- but the wall time is enormous since most of the time only a single thread is running.

Maybe someone will read this and see what's missing, or feel inspired to make their own attempt

What I did
ACML libraries were installed as shown in e.g. http://verahill.blogspot.com.au/2013/05/409-failed-attempt-at-compiling-gamess_10.html

Nwchem was downloaded:
sudo mkdir /opt/nwchem
sudo chown $USER:$USER /opt/nwchem
cd /opt/nwchem
wget http://www.nwchem-sw.org/download.php?f=Nwchem-6.1.1-src.2012-06-27.tar.gz
tar xvf Nwchem-6.1.1-src.2012-06-27.tar.gz
cd nwchem-6.1.1-src/

Next I edited src/config/makefile.h
2363 ifdef OPTIMIZE 2364 FFLAGS = $(FOPTIONS) $(FOPTIMIZE) 2365 CFLAGS = $(COPTIONS) $(COPTIMIZE) -fopenmp 2366 else 2367 # Need FDEBUG after FOPTIONS on SOLARIS to correctly override optimization 2368 FFLAGS = $(FOPTIONS) $(FDEBUG) 2369 CFLAGS = $(COPTIONS) $(CDEBUG) -fopenmp 2370 endif 2371 INCLUDES = -I. $(LIB_INCLUDES) -I$(INCDIR) $(INCPATH) 2372 CPPFLAGS = $(INCLUDES) $(DEFINES) $(LIB_DEFINES) 2373 LDFLAGS = $(LDOPTIONS) -L$(LIBDIR) $(LIBPATH) 2374 LIBS = $(NW_MODULE_LIBS) $(CORE_LIBS) -lgomp 2375
I then built using the following build script:
export LARGE_FILES=TRUE export TCGRSH=/usr/bin/ssh export NWCHEM_TOP=`pwd` export NWCHEM_TARGET=LINUX64 export NWCHEM_MODULES="all" export PYTHONVERSION=2.7 export PYTHONHOME=/usr export BLASOPT="-L/opt/acml/acml5.3.1/gfortran64_fma4_mp_int64/lib -lacml_mp -lpthread" export USE_OPENMP=y export LIBRARY_PATH="$LIBRARY_PATH:/opt/acml/acml5.3.1/gfortran64_fma4_mp_int64/lib" cd $NWCHEM_TOP/src make clean make nwchem_config make FC=gfortran 2> make.err 1>make.log cd $NWCHEM_TOP/contrib export FC=gfortran ./getmem.nwchem
So far so good.

Where it fails
A picture is probably in order:
Note that while this is a short run, it is perfectly representative of what I'm seeing with 'real' jobs too -- I get eight threads auto-spawning (as seen by top), but only one thread is active most of the time.

Basically, most of the time only one core is running at 100% (i.e. showing as 12.5 % here since I have 8 cores), with the other cores occasionally kicking in (the 'spikes').

The wall times is 63 seconds, and the 'cpu time' is 83.1 seconds. Ideally, for a fully parallel run the cpu time should be as close to the wall time multiplied with eight for a shared run like this (but is always smaller).

As a comparison, here's an mpi-enabled binary:
Here all cores are active over most of the (short) run. The cpu time was 9.9 seconds and the wall time 11.8 seconds. For an mpi run the wall time should be as close to the cpu time as possible (but is always larger)

So it's not particularly 'parallel' in the OMP case -- but I don't know why. Maybe nwchem 6.1.1 isn't quite ready for OMP yet? I've noticed that it's one of the areas where the upcoming release is supposed to have been improved.

'profiling' with sar -- how-to
sudo apt-get install syssstat

Edit /etc/default/sysstat:
8 # will be overwritten by debconf! 9 ENABLED="true" 10
sudo service sysstat restart

Before launching the run, set sar to run in another windows and collect data before immediately launching the run you want to monitor in a different window:

sar 1 180 >> run.log
collects data every 1 seconds and repeats it 180 times (i.e. 181 seconds) and stores the data in run.log.

No comments:

Post a Comment