Closed dmageeLANL closed 7 months ago
STREAM questions:
-O2 -fopenmp
as is done here: https://github.com/jeffhammond/STREAM/blob/master/Makefilesrun
commands now include --cpu-bind=core
stream-mpi.exe
executable used? I assumed yes, and the srun
commands now reflect thatSTREAM_ARRAY_SIZE
set to 40,000,000 rather than 40? That appears to give similar results to what is on the page (I used gcc 9.4.0 and openmpi/3.1.6 on Darwin's spr-hbm partition)STREAM changes:
-DSTREAM_ARRAY_SIZE=40000000 -DNTIMES=20
) in the build section along with optimization flags (-O2 -fopenmp
)export OMP_NUM_THREADS=1
to the running section commands--cpu-bind=core
to the srun command./stream
to ./stream-mpi.exe
STREAM_ARRAY_SIZE=40
is now STREAM_ARRAY_SIZE=40000000
OSUMB questions:
Running
section gives enough info, but I wanted to double-check with you.N/A
there for now.OSUMB changes:
Purpose
section.DGEMM questions:
dgemm_omp_fixes.patch
file referenced in the documentation. I am leaving that as is for now.LDFLAGS
was not passed to the make rule and wasn't linking. I did not check Makefile.intel
yet.DGEMM changes:
Example Results
section that mentioned Branson
instead of DGEMM
LDFLAGS
to the make rule and to append to CFLAGS
rather than overwrite CFLAGS
. Did not make any changes to Makefile.intel
make
command in documentation to reference passing the include directory. It is now make CFLAGS=-I<openblas_include_dir>
Done!
Please proofread the microbenchmark docs for STREAM, OSU microbenchmarks and DGEMM to make sure they make sense from a third party POV.