Getting better speed-ups with OPENMP on Linux using ifort

General scientific issues regarding ROMS

Moderators: arango, robertson

Post Reply
Message
Author
lanerolle
Posts: 157
Joined: Mon Apr 28, 2003 5:12 pm
Location: NOAA

Getting better speed-ups with OPENMP on Linux using ifort

#1 Unread post by lanerolle »

I have got the ROMS to run in parallel on a Linux platform (with Fedora 2) using the ifort 8.1 compiler. When doing so, it was necessary for me to change the stacksize on each of the threads (I used 2 processors with each having 2 threads) and I did so using the command : export KMP_STACKSIZE=50m (so that each thread had a 50 MB size). The compilation flags I used in my Makefile were:

CFT = ifort -pc80
CPPFLAGS = -P -DLINUX -D_OPENMP -I$(NETCDF_INCDIR)
CPP = /lib/cpp -traditional
FFLAGS = -i_dynamic -openmp -fpp -O
MDEPFLAGS = --cpp --fext=f90 --file=-

I found that it was absolutely necessary to use the -fpp flag. The resulting executable gave me modest speed-ups over a corresponding serial run (about 30% faster at best when running on 4 threads). I was expecting to get a factor of 2 (100% speed-up) or even better. I found that when I used the flags in Makefile.OMP_ifort (the relevent Makefile which came with the ROMS code package) with FFLAGS = -openmp -fpp -ip -O3 -tpp7 -xW (I had to include -fpp also), the code took a very long time to compile but I got no appreciable increases in speed-ups in the run times.

Does anybody know what additional/alternative compilation flags I can use to get better speed-ups with OPENMP on a Linux platform with ifort?

PS. I have verified that my serial and parallel runs give more or less identical numerical solutions by looking at time histories of the differences in all of the flow variables.

Post Reply