28 October 2012

268. Compiling and testing GAMESS US on debian testing (wheezy)

Update 3: 9 May 2013. Fixed a couple of mistakes e.g. related to mpi. I also switched from ATLAS to acml -- when I build with ATLAS a lot of the example inputs do not converge.

Update 2: Pietro (see posts below) identified some odd behaviour when running test exam44 in which the scf failed to converge. The (temporary) fix for that has been included in the instructions below (change line 1664 in the file 'comp') -- most likely it's only a single file which needs to be compiled with -O0, but it will take a while to identify which one that is. Having to use -O0 on a performance critical piece of software is obviously unfortunate.

I've done this on ROCKS 5.4.3/ Centos 5.6 as well. Be aware that because of the ancient version of gfortran (4.1.2) on ROCKS there will be some limitations:

   Alas, your version of gfortran does not support REAL*16,
   so relativistic integrals cannot use quadruple precision.
   Other than this, everything will work properly.
Other than that, follow the instructions below (including editing lked)

Original post:
Solvation energies using implicit solvation is a tough nut to crack. I like working with NWChem, but there's only one solvation model (COSMO) implemented, it has had a history of giving results which are wildly different (>20 kcal/mol! It's fixed now -- using b3lyp/6-311++g** with the cosmo parameters in that post I got 63.68 kcal/mol for Cl-) from that of other software packages (partly due to a bug which was fixed in 2011), and I'm still not sure how to properly use the COSMO module (is rsolv 0 a reasonable value?). Obviously, my own unfamiliarity with the method is another issue, but that's where the idea of sane defaults come in. So, time to test and compare with other models. Reading Cramer, C. J.; Truhlar, D. G. A Acc. Chem. Res. 2008, 41, 760–768 got me interested in GAMESS US again.

Gaussian is not really an attractive option for me anymore for performance reasons (caveat: as seen by me on my particular systems using precompiled binaries). Free (source code + cost) is obviously also always attractive. Being a linux sort of person also plays into it.

So, here's how to get your cluster set up for gamess US:
1. Go to http://www.msg.chem.iastate.edu/GAMES S/download/register/
Select agree, then pick your version -- in my case
GAMESS version May 1, 2012 R1 for 64 bit (x86_64 compatible) under Linux with gnu compilers

Once you've completed your order you're told you may have to wait for up to a week before your registration is approved, but I got approved in less than 24 hours.

[2. Register for GAMESSPLUS at http://comp.chem.umn.edu/license/form-user.html
Again, it may take a little while to get approved -- in my case it was less than 24 hours. Also, it seems that you don't need a separate GAMESSPLUS anymore]

3. Download gamess-current.tar.gz as per the instructions and put it in /opt/gamess (once you've created the folder)

4. If you're using AMD you're in luck -- set up acml on your system. In my case I put everything in /opt/acml/acml5.2.0

I've had bad luck with ATLAS.

5. Compile
sudo apt-get install build-essential gfortran openmpi-bin libopenmpi-dev libboost-all-dev
sudo mkdir /opt/gamess
sudo chown $USER /opt/gamess
cd /opt/gamess
tar xvf gamess-current.tar.gz 
cd gamess/

You're now ready to autoconfigure.

The lengthy autoconfigure.
 Note that
* the location of your openmpi libs may vary -- the debian libs are put in /usr/bin/openmpi/lib by default, but I'm using my own compiled version which I've put in /opt/openmpi
* gamess is linked against the static libraries by default, so if you compiled atlas as is described elsewhere on this blog, you'll be fine.

This script asks a few questions, depending on your computer system,
to set up compiler names, libraries, message passing libraries,
and so forth.
You can quit at any time by pressing control-C, and then .
Please open a second window by logging into your target machine,
in case this script asks you to 'type' a command to learn something
about your system software situation.  All such extra questions will
use the word 'type' to indicate it is a command for the other window.
After the new window is open, please hit  to go on.

   GAMESS can compile on the following 32 bit or 64 bit machines:
axp64    - Alpha chip, native compiler, running Tru64 or Linux
cray-xt  - Cray's massively parallel system, running CNL
hpux32   - HP PA-RISC chips (old models only), running HP-UX
hpux64   - HP Intel or PA-RISC chips, running HP-UX
ibm32    - IBM (old models only), running AIX
ibm64    - IBM, Power3 chip or newer, running AIX or Linux
ibm64-sp - IBM SP parallel system, running AIX
ibm-bg   - IBM Blue Gene (P or L model), these are 32 bit systems
linux32  - Linux (any 32 bit distribution), for x86 (old systems only)
linux64  - Linux (any 64 bit distribution), for x86_64 or ia64 chips
           AMD/Intel chip Linux machines are sold by many companies
mac32    - Apple Mac, any chip, running OS X 10.4 or older
mac64    - Apple Mac, any chip, running OS X 10.5 or newer
sgi32    - Silicon Graphics Inc., MIPS chip only, running Irix
sgi64    - Silicon Graphics Inc., MIPS chip only, running Irix
sun32    - Sun ultraSPARC chips (old models only), running Solaris
sun64    - Sun ultraSPARC or Opteron chips, running Solaris
win32    - Windows 32-bit (Windows XP, Vista, 7, Compute Cluster, HPC Edition)
win64    - Windows 64-bit (Windows XP, Vista, 7, Compute Cluster, HPC Edition)
winazure - Windows Azure Cloud Platform running Windows 64-bit
    type 'uname -a' to partially clarify your computer's flavor.
please enter your target machine name: linux64

Where is the GAMESS software on your system?
A typical response might be /u1/mike/gamess,
most probably the correct answer is /home/me/tmp/gamess
GAMESS directory? [/opt/gamess] /opt/gamess

Setting up GAMESS compile and link for GMS_TARGET=linux64
GAMESS software is located at GMS_PATH=/home/me/tmp/gamess
Please provide the name of the build locaation.
This may be the same location as the GAMESS directory.
GAMESS build directory? [/opt/gamess] /opt/gamess

Please provide a version number for the GAMESS executable.
This will be used as the middle part of the binary's name,
for example: gamess.00.x

Version? [00] 12r1

Linux offers many choices for FORTRAN compilers, including the GNU
compiler set ('g77' in old versions of Linux, or 'gfortran' in
current versions), which are included for free in Unix distributions.
There are also commercial compilers, namely Intel's 'ifort',
Portland Group's 'pgfortran', and Pathscale's 'pathf90'.  The last
two are not common, and aren't as well tested as the others.
type 'rpm -aq | grep gcc' to check on all GNU compilers, including gcc
type 'which gfortran'  to look for GNU's gfortran (a very good choice),
type 'which g77'       to look for GNU's g77,
type 'which ifort'     to look for Intel's compiler,
type 'which pgfortran' to look for Portland Group's compiler,
type 'which pathf90'   to look for Pathscale's compiler.
Please enter your choice of FORTRAN: gfortran

gfortran is very robust, so this is a wise choice.

Please type 'gfortran -dumpversion' or else 'gfortran -v' to
detect the version number of your gfortran.
This reply should be a string with at least two decimal points,
such as 4.1.2 or 4.6.1, or maybe even 4.4.2-12.
The reply may be labeled as a 'gcc' version,
but it is really your gfortran version.
Please enter only the first decimal place, such as 4.1 or 4.6:  

   Good, the newest gfortran can compile REAL*16 data type.
hit <return> to continue to the math library setup.

Linux distributions do not include a standard math library.
There are several reasonable add-on library choices,
       MKL from Intel           for 32 or 64 bit Linux (very fast)
      ACML from AMD             for 32 or 64 bit Linux (free)
     ATLAS from www.rpmfind.net for 32 or 64 bit Linux (free)
and one very unreasonable option, namely 'none', which will use
some slow FORTRAN routines supplied with GAMESS.  Choosing 'none'
will run MP2 jobs 2x slower, or CCSD(T) jobs 5x slower.
Some typical places (but not the only ones) to find math libraries are
Type 'ls /opt/intel/mkl'                 to look for MKL
Type 'ls /opt/intel/Compiler/mkl'        to look for MKL
Type 'ls /opt/intel/composerxe/mkl'      to look for MKL
Type 'ls -d /opt/acml*'                  to look for ACML
Type 'ls -d /usr/local/acml*'            to look for ACML
Type 'ls /usr/lib64/atlas'               to look for Atlas
Enter your choice of 'mkl' or 'atlas' or 'acml' or 'none': acml

Type 'ls -d /opt/acml*' or 'ls -d /usr/local/acml*'
and note the the full path, which includes a version number.
enter this full pathname: /opt/acml/acml5.2.0
Math library 'acml' will be taken from /opt/acml/acml5.2.0/gfortran64_int64/lib
please hit <return> to compile the GAMESS source code activator
gfortran -o /home/me/tmp/gamess/build/tools/actvte.x actvte.f
unset echo
Source code activator was successfully compiled.
please hit  to set up your network for Linux clusters.

If you have a slow network, like Gigabit Ethernet (GE), or
if you have so few nodes you won't run extensively in parallel, or
if you have no MPI library installed, or
if you want a fail-safe compile/link and easy execution,
     choose 'sockets'
to use good old reliable standard TCP/IP networking.
If you have an expensive but fast network like Infiniband (IB), and
if you have an MPI library correctly installed,
     choose 'mpi'.
communication library ('sockets' or 'mpi')? mpi

The MPI libraries which work well on linux64/Infiniband are
      Intel's MPI (impi)
      SGI's mpt from ProPack, on Altix/ICE systems
Other libraries may work, please see 'readme.ddi' for info.
The choices listed above will compile and link easily,
and are known to run correctly and efficiently.

Enter 'sockets' if you just changed your mind about trying MPI.

Enter MPI library (impi, mvapich2, mpt, sockets): openmpi
MPI can be installed in many places, so let's find openmpi.
The person who installed your MPI can tell you where it really is.
impi     is probably located at a directory like
         include iMPI's version numbers in your reply below.
mvapich2 could be almost anywhere, perhaps some directory like
openmpi  could be almost anywhere, perhaps some directory like
mpt      is probably located at a directory like
Please enter your openmpi's location: /opt/openmpi/1.6

Your configuration for GAMESS compilation is now in
Now, please follow the directions in

I next did this:

cd /opt/gamess/ddi
cd ../

Edit the file 'comp' and change it from

1664       set OPT='-O2'
1664       set OPT='-O0'

or test case exam44.inp in tests/standard will fail due to lack of SCF convergence. (I've tried -O1 as well with no luck)

Continue your compilation:

Running 'compall' reads "install.info" which I include below:
#   compilation configuration for GAMESS
#   generated on beryllium
#   generated at Friday 21 September  08:48:09 EST 2012
setenv GMS_PATH            /opt/gamess
setenv GMS_BUILD_DIR       /opt/gamess
#         machine type
setenv GMS_TARGET          linux64
#         FORTRAN compiler setup
setenv GMS_FORTRAN         gfortran
#         mathematical library setup
setenv GMS_MATHLIB         acml
setenv GMS_MATHLIB_PATH    /opt/acml/acml5.2.0//gfortran64_int64/lib
#         parallel message passing model setup
setenv GMS_DDI_COMM        mpi
setenv GMS_MPI_LIB         openmpi
setenv GMS_MPI_PATH        /opt/openmpi/1.6

Note that you can't change the gfortran version here either -- 4.7 won't be recognised.

Anyway, compilation will take a while -- enough for some coffee and reading.

In the next step you may have problems with openmpi -- lked looks in e.g. /opt/openmpi/1.6/lib64 but you'll probably only have /opt/openmpi/1.6/lib

Edit lked and change
 958             case openmpi:
 959                set MPILIBS="-L$GMS_MPI_PATH/lib64"
 960                set MPILIBS="$MPILIBS -lmpi"
 961                breaksw
 958             case openmpi:
 959                set MPILIBS="-L$GMS_MPI_PATH/lib"
 960                set MPILIBS="$MPILIBS -lmpi -lpthread"
 961                breaksw

Generate the runtime file:
./lked gamess 12r1 >&  lked.log


To compile with openblas:
1. edit install.info
#   compilation configuration for GAMESS
#   generated on tantalum
#   generated at Friday 21 September  14:01:54 EST 2012
setenv GMS_PATH            /opt/gamess
setenv GMS_BUILD_DIR       /opt/gamess
#         machine type
setenv GMS_TARGET          linux64
#         FORTRAN compiler setup
setenv GMS_FORTRAN         gfortran
#         mathematical library setup
setenv GMS_MATHLIB         openblas
setenv GMS_MATHLIB_PATH    /opt/openblas/lib
#         parallel message passing model setup
setenv GMS_DDI_COMM        mpi
setenv GMS_MPI_LIB         openmpi
setenv GMS_MPI_PATH        /opt/openmpi/1.6

2. edit lked
Add lines 462-466 which sets up the openblas switch.

 453       endif
 454       set BLAS=' '
 455       breaksw
 457    case acml:
 458       #     do a static link so that only compile node needs to install ACML
 459       set MATHLIBS="$GMS_MATHLIB_PATH/libacml.a"
 460       set BLAS=' '
 461       breaksw
 462 case openblas:
 463        #     do a static link so that only compile node needs to install openblas
 464        set MATHLIBS="$GMS_MATHLIB_PATH/libopenblas.a"
 465        set BLAS=' '
 466        breaksw
 468    case none:
 469    default:
 470       echo "Warning.  No math library was found, you should install one."
 471       echo "    MP2 calculations speed up about 2x with a math library."
 472       echo "CCSD(T) calculations speed up about 5x with a math library."
 473       set BLAS='blas.o'
 474       set MATHLIBS=' '
 475       breaksw

3. Link
./lked gamess 12r2 >&  lked.log

You now have gamess.12r1.x which uses ATLAS, and gamess.12r2.x which uses openblas.

To run:
The rungms file was a bit too 'clever' for me, so I boiled it down to a file called gmrun which made executable (chmod +X gmrun) and put in /opt/gamess:

set TARGET=mpi
set SCR=$HOME/scratch
set USERSCR=/scratch
set GMSPATH=/opt/gamess
set JOB=$1
set VERNO=$2
set NCPUS=$3

if ( $JOB:r.inp == $JOB ) set JOB=$JOB:r
echo "Copying input file $JOB.inp to your run's scratch directory..."
cp $JOB.inp $SCR/$JOB.F05

setenv INPUT $SCR/$JOB.F05
setenv PUNCH $USERSCR/$JOB.dat
if ( -e $TRAJECT ) rm $TRAJECT
if ( -e  $PUNCH ) rm $PUNCH
if ( -e  $RESTART ) rm $RESTART
source $GMSPATH/gms-files.csh

setenv LD_LIBRARY_PATH /opt/openmpi/1.6/lib:$LD_LIBRARY_PATH
set path= ( /opt/openmpi/1.6/bin $path )
/opt/openmpi/1.6/bin/mpiexec -n $NCPUS $GMSPATH/gamess.$VERNO.x|tee $JOB.out
cp $PUNCH .

Note that I actually do have two scratch directories -- one ~/scratch and one in /scratch. Note that the SCR directory should be local to the node as well as spacious, while USERSCR can be a networked,smaller directory.

Finally do
echo 'export PATH=$PATH:/opt/gamess' >> ~/.bashrc

Navigate to your tests/standard folder where there's a lot of exam*.inp files and do

gmrun exam12 12r1 4

where exam01 (or exam01.inp) is the name of the input file, 12r1 is the version number (that you set above) and 4 is the number of processors/threads .

 POINT   1           X           Y           Z (BOHR)    CHARGE
                 0.000000   -0.000000    0.000000       -0.00 (A.U.)
         DX          DY          DZ         /D/  (DEBYE)
    -0.000000    0.000000   -0.000000    0.000000
 CPU     0: STEP CPU TIME=     0.02 TOTAL CPU TIME=        2.2 (    0.0 MIN)
          IVIB=   0 IATOM=   0 ICOORD=   0 E=      -76.5841347569
 -6.175208802E-40-6.175208802E-40-4.411868660E-07 6.175208802E-40 6.175208802E-40
  4.411868660E-07-1.441225933E-40-1.441225933E-40 1.672333111E-06 1.441225933E-40
 -4.053383177E-34 4.053383177E-34-2.257541709E-15
 CPU     0: STEP CPU TIME=     0.00 TOTAL CPU TIME=        2.2 (    0.0 MIN)
               990473  WORDS OF DYNAMIC MEMORY USED
 DDI: 263624 bytes (0.3 MB / 0 MWords) used by master data server.

 CPU timing information for all processes
 0: 2.160 + 0.44 = 2.204
 1: 2.220 + 0.20 = 2.240
 2: 2.212 + 0.32 = 2.244
 3: 4.240 + 0.04 = 4.244
 4: 4.260 + 0.00 = 4.260
 5: 4.256 + 0.08 = 4.264


Looking at another test case (acetate w/ cosmo) I get the following scaling on a single node as a function of processors:

shmmax issue:
Anyone who has been using nwchem will be familiar with this
 INPUT CARD> $END                                                                           
 DDI Process 0: shmget returned an error.
 Error EINVAL: Attempting to create 160525768 bytes of shared memory.
 Check system limits on the size of SysV shared memory segments.

 The file ~/gamess/ddi/readme.ddi contains information on how to display
 the current SystemV memory settings, and how to increase their sizes.
 Increasing the setting requires the root password, and usually a sytem reboot.

 DDI Process 0: error code 911

The fix is the same. First do

cat /proc/sys/kernel/shmmax

and look at the value. Then set it to the desired value according to this post: http://verahill.blogspot.com.au/2012/04/solution-to-nwchem-shmmax-too-small.html
sudo sysctl -w kernel.shmmax=6269961216

gfortran version issue:
Even though you likely have version 4.7.x of gfortran, pick 4.6 or you will get:

Please type 'gfortran -dumpversion' or else 'gfortran -v' to
detect the version number of your gfortran.
This reply should be a string with at least two decimal points,
such as 4.1.2 or 4.6.1, or maybe even 4.4.2-12.
The reply may be labeled as a 'gcc' version,
but it is really your gfortran version.
Please enter only the first decimal place, such as 4.1 or 4.6:  

The gfortran version number is not recognized.
It should only have one decimal place, such as 4.x

The reason is this (code from config):
      switch ($GMS_GFORTRAN_VERNO)
         case 4.1:
         case 4.2:
         case 4.3:
         case 4.4:
         case 4.5:
            echo "   Alas, your version of gfortran does not support REAL*16,"
            echo "   so relativistic integrals cannot use quadruple precision."
            echo "   Other than this, everything will work properly."
         case 4.6:
            echo "   Good, the newest gfortran can compile REAL*16 data type."
            echo "The gfortran version number is not recognized."
            echo "It should only have one decimal place, such as 4.x"
            exit 4


  1. Hi,

    have you run GAMESS tests after compiling with gfortran 4.7? I compiled using both (a home repackaged version, with x86_64 ss3 optimization, of) atlas, and with OpenBLAS on a Fedora Core 17 machine featuring gcc 4.7 compilers and in both cases I obtained a strange result: 46 of 47 tests were passed but exam44, which doesn't converge, since SCF exhibits an oscillatory behaviour. Searching the web somebody else apparently got the same issue, although the page I found only reported the output of the test summary.
    The same 12r1 version, compiled on a FC16 box with gcc 4.6 and standard atlas, passed all tests.
    Later I'm going to perform other experiments by mixing gcc versions,math libraries and machines. If you are interested, I'll send a more complete report of my tests.

    Pietro Amodeo, ICB-CNR Pozzuoli (NA( - Italy

    1. I tried running test 44, and I observe the same thing as you did in that the energy oscillates and doesn't converge:
      1 -598.7948005066
      2 144.4750515060
      3 140.1905075316
      4 140.1538501955
      5 140.1904325458
      6 51.1643688518
      7 140.1477528877
      8 140.1904315544
      9 140.1538500948
      10 51.1643688561
      11 140.1477528877
      12 140.1904315544
      13 140.1538500948
      14 51.1643688561
      15 140.1477528877
      16 140.1904315544
      17 140.1538500947
      18 51.1643688561
      19 140.1477528877
      20 140.1904315544
      21 140.1538500947
      22 51.1643688561
      23 140.1477528877
      24 140.1904315544
      25 140.1538500947
      26 51.1643688561
      27 140.1477528877
      28 140.1904315544
      29 140.1538500947

    2. And now I've tried it using a version compiled with gcc 4.1.2 -- it converges:
      1 -598.7948005066
      2 -599.9236549126
      3 -599.9656190200
      4 -599.9683370713
      5 -599.9687731854
      6 -599.9687786560
      7 -599.9687798305
      8 -599.9687803241
      9 -599.9687803808
      10 -599.9687803950
      11 -599.9687803940
      12 -599.9687803934

    3. Hi,
      I confirmed that it's a gcc/gfortran 4.7 issue. In particular I compiled GAMESS with:
      1) on a Fedora 17 box with gcc/gfortran 4.7 + ATLAS (rebuilt with gcc 4.7 and repackaged to optimize them on the same box used for calculations. Incidentally, 4.7 required a patch to compile...);
      2) same box and compilers as 1) + OpenBLAS built on the same box with 4.7 compilers;
      3) same box as 1-2), gcc/gfortran 4.6 + same ATLAS as 1;
      4) same box as 1-3), gcc/gfortran 4.6 + OpenBLAS built on the same box with 4.6 compilers;
      5) a different Fedora 16 box, gcc/gfortran 4.6 + standard ATLAS;
      6) same box and compilers as 5) + OpenBLAS built on the same box with 4.6 compilers.

      The strange oscillating behaviour was observed for exam44 when running 1) and 2) executables only, the other combinations were all OK.

      I also made some tests on a real input (a 6-31G* HF geometry optimization) to evaluate performances of all the above versions and the overall behaviour of 1-2). These latter provided reasonable results, almost identical to those obtained with all other executables.
      Performances were quite similar between ATLAS and OpenBLAS on both the architectures I tested (double Xeon boxes with either Core2 or Nehalem processors), but I sadly still have no Sandy Bridge CPU handy...

      An additional note: I had to compile the single-threaded version of OpenBLAS to prevent runtime issues when using > 1/2 of the available cores and performance issues in all cases (maybe advantages can be obtained when running calculation types making a heavier usage of OpenBLAS routines than the simple HF geometry optimization I used to compare performances).


    4. Pietro,
      I think it might be wise to contact Mike Schmidt (http://www.msg.chem.iastate.edu/~mike/) or possibly first post a general enquiry at the independent gamess list at http://groups.google.com/group/gamess

      You have a compelling set of data for what seems to be a bug, or at least unintended behaviour.

      I'm not an expert, so don't know if the answer is here: http://gcc.gnu.org/gcc-4.7/changes.html

      I will recompile gamess in a few days without any optimisation (-O0) and see if the problem persists. Something I read somewhere got me thinking...

      I'll post the outcome here.

    5. Pietro,
      it's the optimisation. I set the optimisation to -O0 in ddi/compddi (lines 433, 699, 707), comp (line 1664) and lked (line 487).

      This time exam44.inp runs perfectly.

  2. Perfect. I like you blog and I add it to my favorite.

  3. Dear lindqvist,

    Do you have experience in running GAMESS (US) with GPU (NVIDIA CUDA)?

    Best regards, Kirill.

    1. Kirill,
      I don't have any experience of using GAMESS(US) with CUDA support since I lack suitable (i.e. dynamically cooled) hardware -- also, I have to admit that I haven't been using gamess much since nwchem has served me well (wish it had a wider range of solvent models though). Have you managed to compile GAMESS with GPU support?

    2. Dear lindqvist,
      Unfortunatelly there's a problem. Compilation successful, but GPU doesn't works, only CPU calculations! I tried to understand my problem, looked hard to run-scripts, but no idea, what's wrong. I wrote a message to you and to GAMESS_forum. Well... hope I'll find the answer... because I'm a student, I'll graduate soon, and I really don't what to do -- this is my diploma, main thing of my research (theme: GPU calculations).

    3. What guide did you follow when you built your GPU-enabled version?

      This one? http://combichem.blogspot.com.au/2011/02/compiling-gamess-with-cuda-gpu-support.html

    4. I just noticed that you had posted a comment on that page.

      I don't think you can use mpi and CUDA/GPU.

      That also means that you shouldn't use mpiexec to run.

      Did you obtain a gamess.gpu.x binary? If you do ldd on it, is it properly linked to the CUDA libs?

      Beyond that, I actually don't know what the best way forward is. I'll post something here if I come across anything.

    5. Another idea: Firefly (i.e. former PC Gamess) may or may not support CUDA/GPU. It's not completely clear from the site:

    6. Dear lindqvist,

      First of all I want to thank you very much for your replies!

      Yes, I saw Firefly before GAMESS. In past there was problem with registration on Firefly_forum, and I left that idea, but now it's become good! Thank you!

      Other side, tonight I noticed special GAMESS version - special for NVIDIA GPU's. It was released recently, as Andrey Asadchev (author) say to me month ago! There are new scripts and new documentation inside, so I will work hard right now!

      Still, I have some points to answer to you.

      * Yes, I was linked GAMESS+CUDA using manual above. But also I'm using official GAMESS manual (in .../gamess/libcchem/aaa.readme.1st).

      * I compiled successfully CPU-version (sockets, mpi - doesn't matter, everything works) and GPU-version. Well, both calculates... but in second version GPU still passive! I cannot understand why. Maybe there's incorrect input file?..

      * In your opinion, why I shouldn't using MPI? GPU-block places in MPI-block in script 'rungms'..

    7. Kirill,
      I shouldn't pretend to know things I don't. From what (little) I understand though MPI can spawn threads on CPU cores, but not GPU. Also, you don't manage the GPU cores individually, but as a device. Again, it's all very fuzzy in my head, and I have no expertise in this.

      See e.g. http://stackoverflow.com/questions/10237443/mpi-vs-gpu-vs-hadoop-what-are-the-major-difference-between-these-three-parallel
      "[..] MPI and CUDA are completely orthogonal to each other."

      The only GPU calcs I've done are described here:

      Finally, you probably can't expect all types of calculations to be able to take advantage of the GPU e.g. gamess might only use it for implicit solvation (PCM, COSMO), but do most of the heavy lifting on the CPU. Again, I'm just speculating.

      That's certainly true for Firefly (says it doesn't use it for DFT)

    8. I've since compiled gamess with GPU support and I think I have a better understanding of what Кирилл Березовский was trying to tell me. Like him, I built according to the instructions in aaa.readme.1st, and like him, while everything is working fine, the GPU is not being used.

      Hence, ignore my replies above, and have a look here instead: