On 20 June 2015 at 20:48, Bert Wesarg <[email protected]> wrote: > Lisandro, > > On 06/20/2015 05:03 AM, Lisandro Dalcin wrote: >> >> Open MPI 1.8.6 was released, and this issue seems to be still there. >> Linux binaries are near 3 times larger: >> https://binstar.org/mpi4py/openmpi/files > > > sorry for the delay. The problem appears with commit > open-mpi/ompi-releases@c540cf8d69edfaa51ce7f487efd418515ac8311f > > In this commit, Jeff (in Cc) removed 'export CFLAGS' from the top-level > configure before calling into sub-configure, and thus the sub-configure > piles up its own CFLAGS, which mostly results in '-g -O2' on GNU systems. > > Jeff, the ball is now in your hands. >
I've just built the 1.8.8 tarball, and the issue is still there, compare the sizes of the osx-64 vs. linux-64 builds: https://anaconda.org/mpi4py/openmpi/files?sort=uploaded >> >> On 8 May 2015 at 06:47, Lisandro Dalcin <[email protected]> wrote: >>> >>> A build of 1.8.4 with just "./configure --prefix=..." produces the >>> following VT libraries with reasonable sizes which are similar to >>> previous releases (the largest one is 1.5 MB): >>> >>> $ cd /home/devel/mpi/openmpi/1.8.4/lib/ && ls -lh *.a >>> -rw-r--r--. 1 dalcinl users 531K May 8 13:42 libopen-trace-format.a >>> -rw-r--r--. 1 dalcinl users 44K May 8 13:42 libotfaux.a >>> -rw-r--r--. 1 dalcinl users 664K May 8 13:42 libvt.a >>> -rw-r--r--. 1 dalcinl users 1.1M May 8 13:42 libvt-hyb.a >>> -rw-r--r--. 1 dalcinl users 970K May 8 13:42 libvt-mpi.a >>> -rw-r--r--. 1 dalcinl users 1.5M May 8 13:42 libvt-mpi-unify.a >>> -rw-r--r--. 1 dalcinl users 655K May 8 13:42 libvt-mt.a >>> -rw-r--r--. 1 dalcinl users 65K May 8 13:42 libvt-pomp.a >>> >>> After cheking the build directory, I confirm VT seems to be being >>> built with optimization and no debug symbols: >>> $ grep CFLAGS >>> /home/dalcinl/Devel/BUILD/openmpi-1.8.4/ompi/contrib/vt/vt/Makefile >>> CFLAGS = -O3 -DNDEBUG -finline-functions -fno-strict-aliasing -pthread >>> CFLAGS_FOR_BUILD = -O3 -DNDEBUG -finline-functions -fno-strict-aliasing >>> -pthread >>> >>> However, after building 1.8.5 exactly the same way (i.e, just >>> ./configure --prefix=...), VT libraries have much larger sizes (the >>> largest one is now 16 MB, i.e. 10 times larger): >>> >>> $ cd /home/devel/mpi/openmpi/1.8.5/lib/ && ls -lh *.a >>> -rw-r--r--. 1 dalcinl users 1.7M May 8 13:42 libopen-trace-format.a >>> -rw-r--r--. 1 dalcinl users 181K May 8 13:42 libotfaux.a >>> -rw-r--r--. 1 dalcinl users 1.9M May 8 13:43 libvt.a >>> -rw-r--r--. 1 dalcinl users 2.9M May 8 13:43 libvt-hyb.a >>> -rw-r--r--. 1 dalcinl users 2.8M May 8 13:43 libvt-mpi.a >>> -rw-r--r--. 1 dalcinl users 16M May 8 13:43 libvt-mpi-unify.a >>> -rw-r--r--. 1 dalcinl users 1.9M May 8 13:43 libvt-mt.a >>> -rw-r--r--. 1 dalcinl users 175K May 8 13:43 libvt-pomp.a >>> >>> and after inspecting CFLAGS in the build directory I get: >>> >>> $ grep CFLAGS >>> /home/dalcinl/Devel/BUILD/openmpi-1.8.5/ompi/contrib/vt/vt/Makefile >>> CFLAGS = -g -O2 >>> CFLAGS_FOR_BUILD = -g -O2 >>> >>> However, plase take a look at the Makefile in one directory up: >>> >>> $ grep CFLAGS >>> /home/dalcinl/Devel/BUILD/openmpi-1.8.5/ompi/contrib/vt/Makefile >>> CFLAGS = -O3 -DNDEBUG -finline-functions -fno-strict-aliasing -pthread >>> CFLAGS_WITHOUT_OPTFLAGS = -DNDEBUG -fno-strict-aliasing -pthread >>> >>> So it seems that in 1.8.5 CFLAGS is not being propagated from >>> ompi/contrib/vt to ompi/contrib/vt/vt >>> >>> Given that the Makefile ompi/contrib/vt looks ok, but the all the >>> others in subdirs under ompi/contrib/vt are not, this smells as a >>> build regression your are likely not aware of. >>> >>> >>> >>> -- >>> Lisandro Dalcin >>> ============ >>> Research Scientist >>> Computer, Electrical and Mathematical Sciences & Engineering (CEMSE) >>> Numerical Porous Media Center (NumPor) >>> King Abdullah University of Science and Technology (KAUST) >>> http://numpor.kaust.edu.sa/ >>> >>> 4700 King Abdullah University of Science and Technology >>> al-Khawarizmi Bldg (Bldg 1), Office # 4332 >>> Thuwal 23955-6900, Kingdom of Saudi Arabia >>> http://www.kaust.edu.sa >>> >>> Office Phone: +966 12 808-0459 >> >> >> >> > > -- > Dipl.-Inf. Bert Wesarg > wiss. Mitarbeiter > > Technische Universität Dresden > Zentrum für Informationsdienste und Hochleistungsrechnen (ZIH) > 01062 Dresden > Tel.: +49 (351) 463-42451 > Fax: +49 (351) 463-37773 > E-Mail: [email protected] > -- Lisandro Dalcin ============ Research Scientist Computer, Electrical and Mathematical Sciences & Engineering (CEMSE) Numerical Porous Media Center (NumPor) King Abdullah University of Science and Technology (KAUST) http://numpor.kaust.edu.sa/ 4700 King Abdullah University of Science and Technology al-Khawarizmi Bldg (Bldg 1), Office # 4332 Thuwal 23955-6900, Kingdom of Saudi Arabia http://www.kaust.edu.sa Office Phone: +966 12 808-0459
