[gmx-developers] PP/ PME splitting in cvs version

[David van der Spoel]

Florian Haberl wrote:
> Hi, 
> 
> i've mailed with Carsten Kutzner, and he said there is a huge performance 
> improvement in cvs version of gromacs, which is reached by PP/ PME node 
> splitting. So i want to try it:
> 

The new code is not perfect yet, not all options work. I suggest you 
start by using a waterbox.
Do also check out the domain-decomposition option.

Don't use this for production yet, at least not without double-checking 
the results.

The compilation error will be fixed at some stage, but it's not crucial 
for mdrun.

Thanks for your benchamrks by the way. Quite impressive. I read that 
Intel are also going to deliver Quad core chips this fall.


> Normal calculation: mpirun -np x mdrun -deffnm test -v 
> works fine with cvs version
> 
> so i tried new flag -npme but i 'm not sure how to use it
> 
> mpirun -np 4 mdrun -deffnm test -v -npme 1 (running on two seperate nodes each 
> dual cpu, so one seperate for pme)
> 
> wont work and produce error
> 
> p2_32362:  p4_error: : 8258
> 2 - MPI_ISEND : Invalid count argument is -724992
> [2]  Aborting program !
> [2] Aborting program!
> rm_l_2_32363: (8.171875) net_send: could not write to fd=5, errno = 32
> p2_32362: (8.171875) net_send: could not write to fd=5, errno = 32
> starting mdrun '?'
> 5000 steps,     10.0 ps.
> 
> p1_32323:  p4_error: net_recv read:  probable EOF on socket: 1
> p3_32401: (5.402344) net_recv failed for fd = 6
> p3_32401:  p4_error: net_recv read, errno = : 104
> p0_32284:  p4_error: net_recv read:  probable EOF on socket: 1
> Killed by signal 2.
> Killed by signal 2.
> p0_32284: (24.269531) net_send: could not write to fd=4, errno = 32
> 
> 
> How can i use this new feature? 
> Is it meant with nodes physical cpu/ core or node? 
> 
> 
> 
> Something won't compile too (cvs version is from today):
> 
> uname -a 
> Linux ccc080 2.6.16.21-0.13-smp #1 SMP Mon Jul 17 17:22:44 UTC 2006 x86_64 
> x86_64 x86_64 GNU/Linux
> mpich version: 1.2.6
> fftw: 3.0.1
> 
> ./configure --enable-mpi --without-x --without-xml --enable-fft=fftw3
> make
> 
> then mv -f ".deps/gmx_rdf.Tpo" ".deps/gmx_rdf.Plo"; else rm -f 
> ".deps/gmx_rdf.Tpo"; exit 1; fi
>  mpicc -DHAVE_CONFIG_H -I. -I. -I../../src -I../../include 
> -DGMXLIBDIR=\"/usr/local/gromacs/share/top\" 
> -I/raid1/haberl/bin/fftw/3.0.1-64/include -O3 -fomit-frame-pointer 
> -finline-functions -Wall -Wno-unused -funroll-all-loops -MT gmx_rama.lo -MD 
> -MP -MF .deps/gmx_rama.Tpo -c gmx_rama.c -o gmx_rama.o
>  mpicc -DHAVE_CONFIG_H -I. -I. -I../../src -I../../include 
> -DGMXLIBDIR=\"/usr/local/gromacs/share/top\" 
> -I/raid1/haberl/bin/fftw/3.0.1-64/include -O3 -fomit-frame-pointer 
> -finline-functions -Wall -Wno-unused -funroll-all-loops -MT gmx_rdf.lo -MD 
> -MP -MF .deps/gmx_rdf.Tpo -c gmx_rdf.c -o gmx_rdf.o
> if /bin/sh ../../libtool --tag=CC --mode=compile mpicc -DHAVE_CONFIG_H -I. -I. 
> -I../../src  -I../../include -DGMXLIBDIR=\"/usr/local/gromacs/share/top\" 
> -I/raid1/haberl/bin/fftw/3.0.1-64/include  -O3 -fomit-frame-pointer 
> -finline-functions -Wall -Wno-unused -funroll-all-loops -MT gmx_rms.lo -MD 
> -MP -MF ".deps/gmx_rms.Tpo" -c -o gmx_rms.lo gmx_rms.c; \
> then mv -f ".deps/gmx_rms.Tpo" ".deps/gmx_rms.Plo"; else rm -f 
> ".deps/gmx_rms.Tpo"; exit 1; fi
> gmx_rdf.c: In function 'do_sq':
> gmx_rdf.c:595: error: 'splinevec' undeclared (first use in this function)
> gmx_rdf.c:595: error: (Each undeclared identifier is reported only once
> gmx_rdf.c:595: error: for each function it appears in.)
> gmx_rdf.c:595: error: 'theta' undeclared (first use in this function)
> gmx_rdf.c:653: error: 'ELECTRONMASS_keV' undeclared (first use in this 
> function)
> gmx_rdf.c:653: error: 'KILO' undeclared (first use in this function)
> gmx_rdf.c:653: error: 'ELECTRONVOLT' undeclared (first use in this function)
> gmx_rdf.c:658: warning: passing argument 1 of 'calc_grid' from incompatible 
> pointer type
> gmx_rdf.c:658: error: incompatible type for argument 2 of 'calc_grid'
> gmx_rdf.c:658: error: incompatible type for argument 3 of 'calc_grid'
> gmx_rdf.c:658: warning: passing argument 6 of 'calc_grid' makes pointer from 
> integer without a cast
> gmx_rdf.c:658: error: too few arguments to function 'calc_grid'
> gmx_rdf.c:665: warning: implicit declaration of function 'init_pme'
> gmx_rdf.c:665: warning: assignment makes pointer from integer without a cast
> gmx_rdf.c:690: warning: implicit declaration of function 'spread_on_grid'
>  mpicc -DHAVE_CONFIG_H -I. -I. -I../../src -I../../include 
> -DGMXLIBDIR=\"/usr/local/gromacs/share/top\" 
> -I/raid1/haberl/bin/fftw/3.0.1-64/include -O3 -fomit-frame-pointer 
> -finline-functions -Wall -Wno-unused -funroll-all-loops -MT gmx_rms.lo -MD 
> -MP -MF .deps/gmx_rms.Tpo -c gmx_rms.c -o gmx_rms.o
> make[3]: *** [gmx_rdf.lo] Error 1
> make[3]: *** Waiting for unfinished jobs....
> gmx_rms.c: In function 'gmx_rms':
> gmx_rms.c:776: warning: ignoring return value of 'fwrite', declared with 
> attribute warn_unused_result
> 
> 
> 
> Greetings,
> 
> Florian
> 


-- 
David.
________________________________________________________________________
David van der Spoel, PhD, Assoc. Prof., Molecular Biophysics group,
Dept. of Cell and Molecular Biology, Uppsala University.
Husargatan 3, Box 596,  	75124 Uppsala, Sweden
phone:	46 18 471 4205		fax: 46 18 511 755
spoel at xray.bmc.uu.se	spoel at gromacs.org   http://folding.bmc.uu.se
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++