RE: [gmx-users] Power6 @ 4.7 Ghz vs Xeon E5472 @ 3.00GHz
Date: Wed, 17 Feb 2010 03:12:57 +1100 From: mark.abra...@anu.edu.au To: gmx-users@gromacs.org Subject: Re: [gmx-users] Power6 @ 4.7 Ghz vs Xeon E5472 @ 3.00GHz On 17/02/10 03:06, pim schravendijk wrote: Dear Gromacs users, I have been installing Gromacs on a Power6 4.7 GHz with Infiniband, GBit-Ethernet before calling configure, I set: export LDFLAGS=-L/gpfs/schraven/fftw3/lib export CPPFLAGS=-I/gpfs/schraven/fftw3/include export CC=mpcc_r -q64 export CFLAGS=-O3 -qstrict -qarch=auto -qtune=auto export CXX=mpCC_r -q64 export CXXFLAGS=-O3 -qstrict -qarch=auto -qtune=auto export F77=mpxlf_r -q64 ### O3 gives warning: The NOSTRICT option (default at OPT(3)) has the potential to alter the semantics of a program. export FFLAGS=-O3 -qstrict -qnoprefetch -qarch=auto -qtune=auto export LIBS=-lmass export FLIBS= -xlf -xlf90 export OBJECT_MODE=64 I run on both machines at 8 cores (so within the board), and find the 8 power6 cores to be 4-5 times slower (!) than the 8 intel xeon cores. Does anyone have similar experience between these machines? Are there some things I should check? It's a bit useless to use the power6 at all with these kind of speeds, but they were already here anyway and amazon's 30 day refund on retour has expired already ;) Hopefully I'm doing something wrong and the situation can still be improved? This issue might have been discussed on the list before, have a search. I suspect the issue is that there are no optimized inner loops for Power6 as there are for x86. You can see what GROMACS is finding and doing in the configure output. Probably GROMACS is finding nothing useful and falling back on generic C versions of such loops. You should experiment with --enable-fortran to try the generic FORTRAN versions. This is partly the issue. The other problem is, if I recall correctly, that someone at IBM told me that the VMX units have been put in the power6 at a late state in the design process and that they are therefore not very fast. Is the factor 4 compared to a Nehalem Xeon? That is roughly the difference I would expect. Berk Mark -- gmx-users mailing listgmx-users@gromacs.org http://lists.gromacs.org/mailman/listinfo/gmx-users Please search the archive at http://www.gromacs.org/search before posting! Please don't post (un)subscribe requests to the list. Use the www interface or send it to gmx-users-requ...@gromacs.org. Can't post? Read http://www.gromacs.org/mailing_lists/users.php _ New Windows 7: Find the right PC for you. Learn more. http://windows.microsoft.com/shop-- gmx-users mailing listgmx-users@gromacs.org http://lists.gromacs.org/mailman/listinfo/gmx-users Please search the archive at http://www.gromacs.org/search before posting! Please don't post (un)subscribe requests to the list. Use the www interface or send it to gmx-users-requ...@gromacs.org. Can't post? Read http://www.gromacs.org/mailing_lists/users.php
RE: [gmx-users] Power6 @ 4.7 Ghz vs Xeon E5472 @ 3.00GHz
Ah, I noticed just now that the Xeon model is in the subject line. So it is not a Nehalem Xeon. Then the difference is somewhat larger than I expected. Maybe using fortran could help a bit. Berk From: g...@hotmail.com To: gmx-users@gromacs.org Subject: RE: [gmx-users] Power6 @ 4.7 Ghz vs Xeon E5472 @ 3.00GHz Date: Wed, 17 Feb 2010 11:47:38 +0100 Date: Wed, 17 Feb 2010 03:12:57 +1100 From: mark.abra...@anu.edu.au To: gmx-users@gromacs.org Subject: Re: [gmx-users] Power6 @ 4.7 Ghz vs Xeon E5472 @ 3.00GHz On 17/02/10 03:06, pim schravendijk wrote: Dear Gromacs users, I have been installing Gromacs on a Power6 4.7 GHz with Infiniband, GBit-Ethernet before calling configure, I set: export LDFLAGS=-L/gpfs/schraven/fftw3/lib export CPPFLAGS=-I/gpfs/schraven/fftw3/include export CC=mpcc_r -q64 export CFLAGS=-O3 -qstrict -qarch=auto -qtune=auto export CXX=mpCC_r -q64 export CXXFLAGS=-O3 -qstrict -qarch=auto -qtune=auto export F77=mpxlf_r -q64 ### O3 gives warning: The NOSTRICT option (default at OPT(3)) has the potential to alter the semantics of a program. export FFLAGS=-O3 -qstrict -qnoprefetch -qarch=auto -qtune=auto export LIBS=-lmass export FLIBS= -xlf -xlf90 export OBJECT_MODE=64 I run on both machines at 8 cores (so within the board), and find the 8 power6 cores to be 4-5 times slower (!) than the 8 intel xeon cores. Does anyone have similar experience between these machines? Are there some things I should check? It's a bit useless to use the power6 at all with these kind of speeds, but they were already here anyway and amazon's 30 day refund on retour has expired already ;) Hopefully I'm doing something wrong and the situation can still be improved? This issue might have been discussed on the list before, have a search. I suspect the issue is that there are no optimized inner loops for Power6 as there are for x86. You can see what GROMACS is finding and doing in the configure output. Probably GROMACS is finding nothing useful and falling back on generic C versions of such loops. You should experiment with --enable-fortran to try the generic FORTRAN versions. This is partly the issue. The other problem is, if I recall correctly, that someone at IBM told me that the VMX units have been put in the power6 at a late state in the design process and that they are therefore not very fast. Is the factor 4 compared to a Nehalem Xeon? That is roughly the difference I would expect. Berk Mark -- gmx-users mailing listgmx-users@gromacs.org http://lists.gromacs.org/mailman/listinfo/gmx-users Please search the archive at http://www.gromacs.org/search before posting! Please don't post (un)subscribe requests to the list. Use the www interface or send it to gmx-users-requ...@gromacs.org. Can't post? Read http://www.gromacs.org/mailing_lists/users.php New Windows 7: Find the right PC for you. Learn more. _ Express yourself instantly with MSN Messenger! Download today it's FREE! http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/-- gmx-users mailing listgmx-users@gromacs.org http://lists.gromacs.org/mailman/listinfo/gmx-users Please search the archive at http://www.gromacs.org/search before posting! Please don't post (un)subscribe requests to the list. Use the www interface or send it to gmx-users-requ...@gromacs.org. Can't post? Read http://www.gromacs.org/mailing_lists/users.php
Re: [gmx-users] Power6 @ 4.7 Ghz vs Xeon E5472 @ 3.00GHz
On 17/02/10 03:06, pim schravendijk wrote: Dear Gromacs users, I have been installing Gromacs on a Power6 4.7 GHz with Infiniband, GBit-Ethernet before calling configure, I set: export LDFLAGS=-L/gpfs/schraven/fftw3/lib export CPPFLAGS=-I/gpfs/schraven/fftw3/include export CC=mpcc_r -q64 export CFLAGS=-O3 -qstrict -qarch=auto -qtune=auto export CXX=mpCC_r -q64 export CXXFLAGS=-O3 -qstrict -qarch=auto -qtune=auto export F77=mpxlf_r -q64 ### O3 gives warning: The NOSTRICT option (default at OPT(3)) has the potential to alter the semantics of a program. export FFLAGS=-O3 -qstrict -qnoprefetch -qarch=auto -qtune=auto export LIBS=-lmass export FLIBS= -xlf -xlf90 export OBJECT_MODE=64 I run on both machines at 8 cores (so within the board), and find the 8 power6 cores to be 4-5 times slower (!) than the 8 intel xeon cores. Does anyone have similar experience between these machines? Are there some things I should check? It's a bit useless to use the power6 at all with these kind of speeds, but they were already here anyway and amazon's 30 day refund on retour has expired already ;) Hopefully I'm doing something wrong and the situation can still be improved? This issue might have been discussed on the list before, have a search. I suspect the issue is that there are no optimized inner loops for Power6 as there are for x86. You can see what GROMACS is finding and doing in the configure output. Probably GROMACS is finding nothing useful and falling back on generic C versions of such loops. You should experiment with --enable-fortran to try the generic FORTRAN versions. Mark -- gmx-users mailing listgmx-users@gromacs.org http://lists.gromacs.org/mailman/listinfo/gmx-users Please search the archive at http://www.gromacs.org/search before posting! Please don't post (un)subscribe requests to the list. Use the www interface or send it to gmx-users-requ...@gromacs.org. Can't post? Read http://www.gromacs.org/mailing_lists/users.php