( 8 Jun 94)
**************************************
* *
* Section 5 - Programmer's Reference *
* *
**************************************
This section describes features of the GAMESS
implementation which are true for all machines. See the
section 'hardware specifics' for information on each
machine type. The contents of this section are:
o Installation overview (sequential mode)
o Files on the distribution tape
o Names of source code modules
o Programming conventions
o Parallel version of GAMESS
philosophy
TCGMSG toolkit
installation process
execution process
load balancing
timing examples
broadcast identifiers
o Disk files used by GAMESS
o Contents of DICTNRY master file
Installation overview
GAMESS will run on a number of different machines
under FORTRAN 77 compilers. However, even given the F77
standard there are still a number of differences between
various machines. For example some machines have 32 bit
word lengths, requiring the use of double precision, while
others have 64 bit words and are used in single precision.
Although there are many types of computers, there is
only one (1) version of GAMESS.
This portability is made possible mainly by keeping
machine dependencies to a minimum (that is, writing in
F77, not vendor specific language extensions). The
unavoidable few statements which do depend on the hardware
are commented out, for example, with "*IBM" in columns
1-4. Before compiling GAMESS on an IBM machine, these
four columns must be replaced by 4 blanks. The process of
turning on a particular machine's specialized code is
dubbed "activation".
A semi-portable FORTRAN 77 program to activate the
desired machine dependent lines is supplied with the
GAMESS package as program ACTVTE. Before compiling ACTVTE
on your machine, use your text editor to activate the very
few machine dependent lines in ACTVTE before compiling it.
Be careful not to change the DATA initialization!
The task of building an executable form of GAMESS is
this:
activate compile link
*.SRC ---> *.FOR ---> *.OBJ ---> *.EXE
source FORTRAN object executable
code code code image
where the intermediate files *.FOR and *.OBJ are discarded
once the executable has been linked. It may seem odd at
first to delete FORTRAN code, but this can always be
reconstructed from the master source code using ACTVTE.
The advantage of maintaining only one master version
is obvious. Whenever any improvements are made, they are
automatically in place for all the currently supported
machines. There is no need to make the same changes in a
plethora of other versions.
The control language needed to activate, compile, and
link GAMESS on your brand of computer is probably present
on the distribution tape. These files should not be used
without some examination and thought on your part, but
should give you a starting point.
There may be some control language procedures for one
computer that cannot be duplicated on another. However,
some general comments apply: Files named COMP will
compile a single module. COMPALL will compile all
modules. LKED will link together an executable image.
RUNGMS will run a GAMESS job, and RUNALL will run all the
example jobs.
The first step in installing GAMESS should be to print
the manual. If you are reading this, you've got that
done! The second step would be to get the source code
activator compiled and linked (note that the activator
must be activated manually before it is compiled). Third,
you should now compile all the source modules (if you have
an IBM, you should also assemble the two provided files).
Fourth, link the program. Finally, run all the short
tests, and very carefully compare the key results shown in
the 'sample input' section against your outputs. These
"correct" results are from a VAX, so there may be very
tiny (last digit) precision differences for other
machines. That's it!
Before starting the installation, you should read the
pages decribing your computer in the 'Hardware Specifics'
section of the manual. There may be special instructions
for your machine.
Files for GAMESS
*.DOC The files you are reading now. You
should print these on 8.5 by 11 inch
white paper, using column one as
carriage control. Double sided, 3
hole, 10 pitch laser output is best!
*.SRC source code for each module
*.ASM IBM mainframe assembler source
*.C C code used by some UNIX systems.
EXAM*.INP 21 short test jobs (see TESTS.DOC).
BENCH*.INP 13 longer test jobs.
These are files related to some utility programs:
ACTVTE.CODE Source code activator. Note that you
must use a text editor to MANUALLY
activate this program before using it.
MBLDR.* model builder (internal to Cartesian)
CARTIC.* Cartesian to internal coordinates
CLENMO.* cleans up $VEC groups
There are files related to X windows graphics.
See the file INTRO.MAN for their names.
The remaining files are command language for the
various machines.
*.COM VAX command language. PROBE is especially
useful for persons learning GAMESS.
*.MVS IBM command language for MVS (dreaded JCL).
*.CMS IBM command language for CMS. These should
be copied to filetype EXEC.
*.CSH UNIX C shell command language. These should
have the "extension" omitted, and have their
mode changed to executable.
Names of source code modules
The source code for GAMESS is divided into a number
of sections, called modules, each of which does related
things, and is a handy size to edit. The following is a
list of the different modules, what they do, and notes on
their machine dependencies.
machine
module description dependency
------- ------------------------- ----------
BASECP SBK and HW valence basis sets
BASEXT DH, MC, 6-311G extended basis sets
BASHUZ Huzinaga MINI/MIDI basis sets to Xe
BASHZ2 Huzinaga MINI/MIDI basis sets Cs-Rn
BASN21 N-21G basis sets
BASN31 N-31G basis sets
BASSTO STO-NG basis sets
BLAS level 1 basic linear algebra subprograms
CPHF coupled perturbed Hartree-Fock 1
CPROHF open shell/TCSCF CPHF 1
ECP pseudopotential integrals
ECPHW Hay/Wadt effective core potentials
ECPLIB initialization code for ECP
ECPSBK Stevens/Basch/Krauss/Jasien/Cundari ECPs
EIGEN Givens-Householder, Jacobi diagonalization
EFDMY dummy source file
FFIELD finite field polarizabilities
FRFMT free format input scanner
GAMESS main program, single point energy
and energy gradient drivers, misc.
GRD1 one electron gradient integrals
GRD2A two electron gradient integrals 1
GRD2B " " " "
GRD2C " " " "
GRD2D " " " "
GRD2E " " " "
GRD2F " " " "
GUESS initial orbital guess
GUGDGA Davidson CI diagonalization 1
GUGDGB " " " 1
GUGDM 1 particle density matrix
GUGDM2 2 particle density matrix 1
GUGDRT distinct row table generation
GUGEM GUGA method energy matrix formation 1
GUGSRT sort transformed integrals 1
GVB generalized valence bond HF-SCF 1
(continued...)
machine
module description dependency
------- ------------------------- ----------
HESS hessian computation drivers
HSS1A one electron hessian integrals
HSS1B " " " "
HSS2A two electron hessian integrals 1
HSS2B " " " "
INPUTA read geometry, basis, symmetry, etc.
INPUTB " " " "
INPUTC " " " "
INT1 one electron integrals
INT2A two electron integrals 1
INT2B " " "
INT2C roots for Rys polynomials
IOLIB input/output routines,etc. 2
LAGRAN CI Lagrangian matrix 1
LOCAL various localization methods 1
MCSCF second order MCSCF calculation 1
MCTWO two electron terms for MCSCF 1
MP2 2nd order Moller-Plesset 1
MPCDAT MOPAC parameterization
MPCGRD MOPAC gradient
MPCINT MOPAC integrals
MPCMOL MOPAC molecule setup
MPCMSC miscellaneous MOPAC routines
MTHLIB printout, matrix math utilities 1
NAMEIO namelist I/O simulator
ORDINT sort atomic integrals 1
PARLEY communicate to other programs
PRPEL electrostatic properties
PRPLIB miscellaneous properties
PRPPOP population properties
RHFUHF RHF, UHF, and ROHF HF-SCF 1
RXNCRD intrinsic reaction coordinate
SCFLIB HF-SCF utility routines, DIIS code
SCRF self consistent reaction field
SPNORB 1 e- spin-orbit coupling terms
STATPT geometry and transition state finder
STUB small version dummy routines
SYMORB orbital symmetry assignment
SYMSLC " " " 1
TCGSTB stub routines to link a serial GAMESS
TRANS partial integral transformation 1
TRFDM2 backtransform 2 e- density matrix 1
TRNSTN CI transition moments
TRUDGE nongradient optimization
(continued...)
machine
module description dependency
------- ------------------------- ----------
UNPORT unportable, nasty code 3,4,5,6,7,8
VECTOR vectorized version routines 9
VIBANL normal coordinate analysis
ZMATRX internal coordinates
Ordinarily, you will not use STUB.SRC, which is linked
only if your system has a very small amount of physical
memory.
In addition, the IBM mainframe version uses the following
assembler language routines: ZDATE.ASM, ZTIME.ASM.
UNIX versions may use the C code: ZMIPS.C, ZUNIX.C.
The machine dependencies noted above are:
1) packing/unpacking 2) OPEN/CLOSE statments
3) machine specification 4) fix total dynamic memory
5) subroutine walkback 6) error handling calls
7) timing calls 8) LOGAND function
9) vector library calls
Programming Conventions
The following "rules" should be adhered
to in making changes in GAMESS. These
rules are important in maintaining
portability, and should be strictly
adhered to.
Rule 1. If there is a way to do it that works on all
computers, do it that way. Commenting out statements for
the different types of computers should be your last
resort. If it is necessary to add lines specific to your
computer, PUT IN CODE FOR ALL OTHER SUPPORTED MACHINES.
Even if you don't have access to all the types of
supported hardware, you can look at the other machine
specific examples found in GAMESS, or ask for help from
someone who does understand the various machines. If a
module does not already contain some machine specific
statements (see the above list) be especially reluctant to
introduce dependencies.
Rule 2. a) Use IMPLICIT DOUBLE PRECISION(A-H,O-Z)
specification statements throughout. b) All floating
point constants should be entered as if they were in
double precision. The constants should contain a decimal
point and a signed two digit exponent. A legal constant
is 1.234D-02. Illegal examples are 1D+00, 5.0E+00, and
3.0D-2. c) Double precision BLAS names are used
throughout, for example DDOT instead of SDOT.
The source code activator ACTVTE will
automatically convert these double
precision constructs into the correct
single precision expressions for machines
that have 64 rather than 32 bit words.
Rule 3. FORTRAN 77 allows the use of generic
functions. Thus the routine SQRT should be used in place
of DSQRT, as this will automatically be given the correct
precision by the compilers. Use ABS, COS, INT, etc. Your
compiler manual will tell you all the generic names.
Rule 4. Every routine in GAMESS begins with a card
containing the name of the module and the routine. An
example is "C*MODULE xxxxxx *DECK yyyyyy". The second
star is in column 18. Here, xxxxxx is the name of the
module, and yyyyyy is the name of the routine.
Furthermore, the individual decks yyyyyy are stored in
alphabetical order. This rule is designed to make it
easier for a person completely unfamiliar with GAMESS to
find routines. The trade off for this is that the driver
for a particular module is often found somewhere in the
middle of that module.
Rule 5. Whenever a change is made to a module, this
should be recorded at the top of the module. The
information required is the date, initials of the person
making the change, and a terse summary of the change.
Rule 6. No lower case characters, no more than 6
letter variable names, no imbedded tabs, statements must
lie between columns 7 and 72, etc. In other words, old
style syntax.
* * *
The next few "rules" are not adhered to
in all sections of GAMESS. Nonetheless
they should be followed as much as
possible, whether you are writing new
code, or modifying an old section.
Rule 7. Stick to the FORTRAN naming convention for
integer (I-N) and floating point variables (A-H,O-Z). If
you've ever worked with a program that didn't obey this,
you'll understand why.
Rule 8. Always use a dynamic memory allocation
routine that calls the real routine. A good name for the
memory routine is to replace the last letter of the real
routine with the letter M for memory.
Rule 9. All the usual good programming techniques,
such as indented DO loops ending on CONTINUEs,
IF-THEN-ELSE where this is clearer, 3 digit statement
labels in ascending order, no three branch GO TO's,
descriptive variable names, 4 digit FORMATs, etc, etc.
The next set of rules relates to coding
practices which are necessary for the
parallel version of GAMESS to function
sensibly. They must be followed without
exception!
Rule 10. All open, rewind, and close operations on
sequential files must be performed with the subroutines
SEQOPN, SEQREW, and SEQCLO respectively. You can find
these routines in IOLIB, they are easy to use.
Rule 11. All READ and WRITE statements for the
formatted files 5, 6, 7 (variables IR, IW, IP, or named
files INPUT, OUTPUT, PUNCH) must be performed only by the
master task. Therefore, these statements must be enclosed
in "IF (MASWRK) THEN" clauses. The MASWRK variable is
found in the /PAR/ common block, and is true on the master
process only. This avoids duplicate output from slave
processes. At the present time, all other disk files in
GAMESS also obey this rule.
Rule 12. All error termination is done by means of
"CALL ABRT" rather than a STOP statement. Since this
subroutine never returns, it is OK to follow it with a
STOP statement, as compilers may not be happy without a
STOP as the final executable statment in a routine.
Parallel version of GAMESS
Under the auspices of a joint ARPA and Air Force
project, GAMESS has begun the arduous journey toward
parallelization. Currently, nearly all ab initio
methods run in parallel, although many of these still
have a step or two running sequentially only. Only MP2
for UHF/ROHF has no parallel method coded. In addition,
MOPAC runs can be run on one node only. More information
about the parallel implementation is given below, after the
directions for installation and execution.
If a parallel linked version of GAMESS is run on only
one node, it behaves as if it is a sequential version, and
the full functionality of the program is available to you.
* * *
The two major philosophies for distributed memory MIMD
(multiple instruction on multiple data) parallel programs
are
1) Have a master program running on one node do all of
the work, except that smaller slave programs running
on the other nodes are called to do their part of the
compute intensive tasks, or
2) Have a single program duplicate all work except for
compute intensive code where each node does only its
separate piece of the work (SPMD, which means single
program, multiple data).
We have chosen to implement the SPMD philosophy in
GAMESS for several reasons. The first of these is that
only one code is required (not master and slave codes).
Therefore, two separate GAMESS codes do not need to be
maintained. The second reason is also related to
maintainance. GAMESS is constantly evolving as new code
is incorporated into it. The parallel calls are "hidden"
at the lowest possible subroutine levels to allow
programmers to add their code with a minimum of extra
effort to parallelize it. Therefore, new algorithms or
methods are available to all nodes. The final reason
given here is that duplication of computation generally
cuts down on communication.
The only portion of the master/slave concept to
survive in GAMESS is that the first process (node 0)
handles reading all input lines and does all print out
and PUNCH file output, as well as all I/O to the DICTNRY
master file. In this sense node 0 is a "master". A
reminder here to all programmers: you should STRICTLY
obey the rules for programming laid out in the Programming
Conventions Section of this manual; especially the ones
involving MASWRK in I/O statements!
* * *
Several tools are available for parallelization of
codes. We have chosen to use the parallelization tool
TCGMSG from Robert Harrison, now at Pacific Northwest
Laboratory. This message passing toolkit has been ported
to many UNIX machines and was written specifically for
computational chemistry. It works on distributed memory
MIMD systems, on Ethernetworks of ordinary workstations,
and on shared memory parallel computers. Thus TCGMSG
allows one to run parallel GAMESS on a fairly wide
assortment of hardware.
Be sure to note that TCGMSG does support communication
between Ethernet workstations of different brands and/or
speeds. For example, we have been able to run on a 3 node
parallel system built from a DECstation, a SGI Iris, and a
IBM RS/6000! (see XDR in $SYSTEM.) It is also useful to
note that your Ethernet parallel system does not have to
contain a power of two number of nodes.
TCGMSG uses the best interprocess communication
available on the hardware being used. For a Ethernetwork
of workstations, this means that TCP/IP sockets are used
for communication. In turn, this means it is extremely
unlikely you will be able to include non-Unix systems in a
Ethernet parallel system.
* * *
If you are trying to run on a genuine parallel system
on which TCGMSG does not work, you may still be in luck.
The "stubs" TCGSTB.SRC can be used to translate from the
TCGSMG calls sprinkled throughout GAMESS to some other
message passing language. For example, we are able to
use GAMESS on the IBM SP1, Intel Paragon, and Thinking
Machines CM-5 in this way, so there is no need to install
TCGMSG to run GAMESS on these systems.
* * *
Our experience with parallel GAMESS is that it is quite
robust in production runs. In other words, most of the
grief comes during the installation phase! TCGMSG will
install and execute without any special priviledges.
The first step in getting GAMESS to run in parallel is
to link GAMESS in sequential mode, against the object file
from TCGSTB.SRC, and ensure that the program is working
correctly in sequential mode.
Next, obtain a copy of the TCGMSG toolkit. This is
available by anonymous ftp from ftp.tcg.anl.gov. Go to
the directory /pub/tcgmsg and, using binary mode, transfer
the file tcgmsg.4.04.tar.Z. (or higher version)
Unpack this file with 'uncompress' and 'tar -xvf'.
The only modification we make to TCGMSG before compiling
it is to remove all -DEVENTLOG flags from the prototype
file tcgmsg/ipcv4.0/Makefile.proto. Then, use the makefile
provided to build the TCGMSG library
chdir ~/tcgmsg
make all MACHINE=IBMNOEXT
If your machine is not a IBM RS/6000, substitute the name
of your machine instead. At this point you should try the
simple "hello" example,
chdir ipcv4.0
parallel hello
to make sure TCGMSG is working.
Finally, link GAMESS against the libtcgmsg.a library
instead of tcgstb.o to produce a parallel executable for
GAMESS. It is not necessary to recompile to accomplish
this. Instead just change the 'lked' script and relink.
* * *
Execute GAMESS by modifying the 'pargms' script to
invoke the TCGMSG program 'parallel', according to the
directions within that script. You also must create a
'gamess.p' file in your home directory,such as
# user, host, nproc, executable, workdir
theresa si.fi.ameslab.gov 1
/u/theresa/gamess/gamess.01.x /scr/theresa
windus ge.fi.ameslab.gov 1
/u/windus/gamess/gamess.01.x /wrk/windus
The fields in each line are: username on that
machine, hostname of that machine, number of processes to
be run on that machine, full file name of the GAMESS
executable on that machine, and working directory on that
machine. Comments begin with a # character. Although
TCGMSG allows long lines to continue on to a new line,
as shown above, you should not do this. The execution
script provided with GAMESS will automatically delete
work files established in the temporary directories, but
only if this script gives all host info on a single line.
A detailed explanation of each field follows:
The first hostname given must be the name of the
machine on which you run the 'pargms' script. This
script defines environment variables specifying the
location of the input and output files. The environment
is not passed to other nodes by TCGMSG's "parallel"
program, so the master process (node 0) running "pargms"
**must** be the first line of your gamess.p file.
The hostname may need to be the shortened form, rather
than the full dotted name, especially on SGI and Unicos.
In general, the correct choice is whatever the response
to executing the command "hostname" is.
The processes on other workstations are generated
by use of the Unix 'rsh' command. This means that you
must set up a .rhosts file in your home directory on each
node on which you intend to run in parallel. This file
validates login permission between the various machines,
by mapping your accounts on one machine onto the others.
For example, the following .rhosts file might be in
Theresa's home directory on both systems,
si.fi.ameslab.gov theresa
ge.fi.ameslab.gov windus
You can test this by getting 'rsh' to work, by a command
such as this, (from si.fi.ameslab.gov)
rsh ge.fi.ameslab.gov -l windus 'df'
and then also try it in the reverse direction too.
Note that the number of processes to be started on a
given machine is ordinarily one. The only exception
is if you are running on a multiCPU box, with a common
memory. In this case, gamess.p should contain just one
line, starting n processes on your n CPUs. This will
used shared memory communications rather than sockets to
pass messages, and is more efficient.
The executable file does not have to be duplicated
on every node, although as shown in the example it can
be. If you have a homogenous Ethernet system, and there
is a single file server, the executable can be stored
on this server to be read by the other nodes by NFS.
Of course, if you have a heterogenous network, you must
build a separate executable for each different brand of
computer you have.
At present GAMESS may write various binary files to the
working directory, depending on what kind of run you are
doing. In fact, the only type of run which will not open
files on the other nodes is a direct SCF, non-analytic
hessian job. Any core dump your job might produce will end
up in this work directory as well.
* * *
We have provided you with a script named 'seqgms'
which will run a parallel-linked version of GAMESS using
only one process on your current machine. Seqgms will
automatically build a single node .p file. Using this
script means you need to keep only a parallel-linked
GAMESS executable, and yet you still retain access to the
parts of GAMESS that do not yet run in parallel.
* * *
We turn now to a description of the way each major
parallel section of GAMESS is implemented, and give
some suggestions for efficient usage.
* * *
The HF wavefunctions can be evaluated in parallel
using either conventional disk storage of the integrals,
or via direct recomputation of the integrals. Assuming
the I/O speed of your system is good, direct SCF is
*always* slower than disk storage. But, a direct SCF
might be faster if your nodes access disk via NFS
over the Ethernet, or if you are running on a Intel
or CM-5 machine. But, if you are running on Ethernetted
workstations which have large local disks on each one,
then conventional disk based SCF is probably fastest.
When you run a disk based SCF in parallel, the
integral files are opened in the work directory which
you defined in your gamess.p file. Only the subset
of integrals computed by each node are stored on that
node's disk space. This lets you store integral files
(in pieces) that are larger than will fit on any one
of your computers.
You may wish to experiment with both options, so
that you learn which is fastest on your hardware setup.
* * *
One of the most important issues in parallelization is
load balancing. Currently, GAMESS has two algorithms
available for load balancing of the two-electron integrals
and gradients. The first is a simple inner loop algorithm
(BALTYP=LOOP). The work of the inner most loop is split
up so the next processor works on the next loop occurence.
If all processors are of the same speed and none of the
processors is dedicated to other work (for example, an
Intel), this is the most effective load balance technique.
The second method is designed for systems where the
processor speeds may not be the same, or where some of the
processors are doing other work (such as a system of equal
workstations in which one of them might be doing other
things). In this technique, as soon as a processor
finishes its previous task, it takes the next task on the
list of work to be done. Thus, a faster node can take
more of the work, allowing all nodes to finish the run at
the same time. This method is implemented throughout most
of GAMESS (see BALTYP=NXTVAL in $SYSTEM). It requires
some extra work coordinating which tasks have been done
already, so NXTVAL adds a communication penalty of about
5% to the total run time.
All integral sections (meaning the ordinary integrals,
gradient integrals, and hessian integrals) have both LOOP
and NXTVAL balancing implemented. Thus all of a HF level
run involving only the energy and gradient has both load
balancing techniques. Analytic HF hessians also have both
balancing techniques for the integral transformation step.
The parallel CI/MCSCF program also contains both
balancing algorithms, except that for technical reasons
MCSCF gradient computation will internally switch to the
LOOP balancing method for that step only.
The parallel MP2 program uses only LOOP balancing
during the MP2, although it will use either balancing
method during the preliminary SCF.
The IBM SP1, Intel, CM-5 always use LOOP balancing,
ignoring your input BALTYP.
* * *
You can find performance numbers for conventional and
direct SCF, as well as gradient evaluation in the paper
M.W.Schmidt, et al., J.Comput.Chem. 14, 1347-1363(1993).
Data for the MCSCF program is not yet published, so
we will include one example here. The data are from
IBM RS/6000 model 350s, connected by Ethernet, using
LOOP balancing. CPU times to do one iteration are in
seconds:
# CPUs= 1 2 3 4 5
- - - - -
MO guess 3.8 4.1 5.3 5.2 5.1
AO ints 391.9 392.0 391.5 391.0 391.0
DRT 0.5 0.5 0.6 0.6 0.6
transf 1539.1 764.5 616.2 461.0 304.7
CI 1.4 0.9 0.8 0.7 0.6
DM2 0.1 0.1 0.2 0.2 0.2
Lag+Hess 16.6 20.6 26.8 26.3 25.3
NR 25.2 27.0 25.4 25.4 25.4
The first three steps precede the MCSCF iterations
and are not parallel. The integral transformation, and
generation of the CI Hamiltonian and its diagonalization
are reasonably parallel (the above example has a trivial
number of CSFs). Large global sums overwhelm the
parallel construction of the Lagrangian and orbital
Hessian. At present the Newton-Raphson orbital change
is running sequentially. Not shown: The backtransform
of the DM2 to the AO basis to set up gradient calculation
runs close to sequentially, but the gradient integral
computation is perfectly parallel.
A CI/MCSCF job will open several disk files on each
node. For example, if the integral transformation
is not being run in direct mode (see DIRTRF in $TRANS)
then each node will compute and store a full copy of the
AO integrals. Each node will store a subset of the
transformed integrals, the CI Hamiltonian, and the density
matrix. The algorithm thus demands not only disk storage
on each node, but also a reasonable I/O bandwidth. We
have not had the opportunity to run the MCSCF code on an
Intel Paragon or a Thinking Machines CM-5, which are not
known for their I/O capacity. Similar comments apply to
analytic hessians: you must have disk storage on each
node, reachable at a reasonable bandwidth.
The integral transformation just described is also
used to drive both parallel analytic hessians and energy
localizations. Thus the scalability of parallel hessians
is much better than described by T.L.Windus, M.W.Schmidt,
M.S.Gordon, Chem.Phys.Lett., 216, 375-379(1993), in that
all steps but the coupled Hartree-Fock are now parallel.
The closed shell parallel MP2 computation is adapted
from Michel Dupuis' implementation. In contrast to the
usual transformation, the specialized MP2 transformation
has the AO integrals actually distributed over each node,
instead of being replicated on each. Obviously this uses
much less disk storage! However, since each node needs to
work with the full AO integral list, the subset stored on
each node must be broadcast to all other nodes. Thus the
savings in disk storage comes at the expense of substantial
extra communication. Initial tests show that the extra
communications prevent the MP2 code from scaling very well
when Ethernet (a rather slow communication channel) is used
to tie together workstations. We do not yet have any
information about the code's performance on a machine such
as the IBM SP1 with fast communications.
To summarize, the normal transformation (used by CI,
MCSCF, energy localization, analytic hessians) must store
duplicate lists of AO integrals on each node, but has almost
no internode communication, and thus scales well on even
low speed networks. The MP2 transformation stores only a
subset of AO integrals on each node, but requires high speed
communications to send each integral to all nodes. Time
will tell us which method is wiser. Note that both of the
transformations distribute the memory needs, which can be
substantial on one node, over all the nodes. Both methods
must evaluate the full AO integral list if a direct integral
transformation is done, during each pass. Being able to
distribute the memory (i.e. the passes) over each nodes means
that direct transformations make more sense in parallel than
in sequential mode. In fact, each node may only need to
compute the AO integrals once, instead of many times if the
job is run on one node. Note that both direct transforms
compute the full AO integral list on each node during each
pass, a blatant sequential bottleneck.
All types of ab initio runs (except for UHF/ROHF MP2
energies) should now run in parallel. However, only the
code for HF energies and gradients is mature, so several
sequential bottlenecks remain. The following steps of a
parallel run will be conducted sequentially by the master:
MCSCF: solution of Newton-Raphson equations
analytic hessians: the coupled Hartree-Fock
energy localizations: the orbital localization step
transition moments/spin-orbit: the final property step
However, all other steps (such as the evaluation of the
underlying wavefunction) do speed up in parallel. Other
steps which do not scale well, although they do speed up
slightly are:
HF: solution of SCF equations
MCSCF/CI: Hamiltonian and 2 body density generation
MCSCF: 2 body density back transformation
Future versions of GAMESS will address these bottlenecks.
In the meantime, some trial and error will teach you how
many nodes can effectively contribute to any particular
type of run.
One example, using the same RS/6000-350 machines and
molecule (bench12.inp converted to runtyp=hessian, with
2,000,000 words, with baltyp=loop) gives the following
replacement for Table 1 of the Chem.Phys.Lett. 216,
375-379(1993) paper:
p= 1 2 3
--- --- ---
setup 0.57 0.69 0.73
1e- ints 1.10 0.87 0.88
huckel guess 15.77 15.74 16.17
2e- ints 111.19 55.34 37.42
RHF cycles 223.13 103.26 79.44
properties 2.23 2.46 2.63
2e- ints -- 111.28 110.97
transformation 1113.67 552.38 381.09
1e- hess ints 28.20 16.46 14.63
2e- hess ints 3322.92 1668.86 1113.37
CPHF 1438.66 1433.34 1477.32
------- ------- -------
total CPU 6258.01 3961.34 3235.27
total wall 8623(73%) 5977(66%) 5136(63%)
so you can see the CPHF is currently a hindrance to full
scalability of the analytic hessian program.
List of parallel broadcast numbers
GAMESS uses TCGMSG calls to pass messages between the
parallel processes. Every message is identified by a
unique number, hence the following list of how the numbers
are used at present. If you need to add to these, look at
the existing code and use the following numbers as
guidelines to make your decision. All broadcast numbers
must be between 1 and 32767.
20 : Parallel timing
100 - 199 : DICTNRY file reads
200 - 204 : Restart info from the DICTNRY file
210 - 214 : Pread
220 - 224 : PKread
225 : RAread
230 : SQread
250 - 265 : Nameio
275 - 310 : Free format
325 - 329 : $PROP group input
350 - 354 : $VEC group input
400 - 424 : $GRAD group input
425 - 449 : $HESS group input
450 - 474 : $DIPDR group input
475 - 499 : $VIB group input
500 - 599 : matrix utility routines
800 - 830 : Orbital symmetry
900 : ECP 1e- integrals
910 : 1e- integrals
920 - 975 : EF and SCRF integrals
980 - 999 : property integrals
1000 - 1025 : SCF wavefunctions
1050 : Coulomb integrals
1200 - 1215 : MP2
1300 : localization
1500 : One-electron gradients
1505 - 1599 : EF and SCRF gradients
1600 - 1602 : Two-electron gradients
1605 - 1615 : One-electron hessians
1650 - 1665 : Two-electron hessians
1700 : integral transformation
1800 : GUGA sorting
1850 - 1865 : GUGA CI diagonalization
1900 - 1905 : GUGA DM2 generation
2000 - 2010 : MCSCF
2100 - 2120 : coupled perturbed HF
Disk files used by GAMESS
unit name contents
---- ---- --------
4 IRCDATA archive results punched by IRC runs, and
restart data for numerical HESSIAN runs.
5 INPUT Namelist input file. This MUST be a disk
file, as GAMESS rewinds this file often.
6 OUTPUT Print output (FT06F001 on IBM mainframes)
If not defined, UNIX systems will use the
standard output for this printout.
7 PUNCH Punch output. A copy of the $DATA deck,
orbitals for every geometry calculated,
hessian matrix, normal modes from FORCE,
properties output, IRC restart data, etc.
8 AOINTS Two e- integrals in AO basis
9 MOINTS Two e- integrals in MO basis
10 DICTNRY Master dictionary, for contents see below.
11 DRTFILE Distinct row table file for -CI- or -MCSCF-
12 CIVECTR Eigenvector file for -CI- or -MCSCF-
13 NTNFMLA Newton-Raphson formula tape for -MCSCF-
14 CIINTS Sorted integrals for -CI- or -MCSCF-
15 WORK15 GUGA loops for diagonal elements;
ordered second order density matrix;
scratch storage during Davidson diag
16 WORK16 GUGA loops for off diagonal elements;
unordered second order density matrix;
2nd order density in AO basis
17 CSFSAVE CSF data for transition moments, SOC
18 FOCKDER derivative Fock matrices analytic hess
20 DASORT Sort file for -MCSCF- or -CI-;
also used by HF's DIIS method
23 JKFILE J and K "Fock" matrices for -GVB-
24 ORDINT sorted AO integrals
25 EFPIND effective fragment data
Contents of the direct access file 'DICTNRY'
1. Atomic coordinates
2. various energy quantities in /ENRGYS/
3. Gradient vector
4. Hessian (force constant) matrix
int 5. ISO - symmetry operations for shells
int 6. ISOC - symmetry operations for centers (atoms)
7. PTR - symmetry transformation for p orbitals
8. DTR - symmetry transformation for d orbitals
9. not used, reserved for FTR
10. not used, reserved for GTR
11. Bare nucleus Hamiltonian integrals
12. Overlap integrals
13. Kinetic energy integrals
14. Alpha Fock matrix (current)
15. Alpha orbitals
16. Alpha density matrix
17. Alpha energies or occupation numbers
18. Beta Fock matrix (current)
19. Beta orbitals
20. Beta density matrix
21. Beta energies or occupation numbers
22. Error function extrapolation table
23. Old alpha Fock matrix
24. Older alpha Fock matrix
25. Oldest alpha Fock matrix
26. Old beta Fock matrix
27. Older beta Fock matrix
28. Oldest beta Fock matrix
29. Vib 0 gradient for FORCE runs
30. Vib 0 alpha orbitals in FORCE
31. Vib 0 beta orbitals in FORCE
32. Vib 0 alpha density matrix in FORCE
33. Vib 0 beta density matrix in FORCE
34. dipole derivative tensor in FORCE.
35. frozen core Fock operator
36. Lagrangian multipliers
37. floating point part of common block /OPTGRD/
int 38. integer part of common block /OPTGRD/
39. ZMAT of input internal coords
int 40. IZMAT of input internal coords
41. B matrix of redundant internal coords
42. not used.
43. Force constant matrix in internal coordinates.
44. SALC transformation
45. symmetry adapted Q matrix
46. S matrix for symmetry coordinates
47. ZMAT for symmetry internal coords
int 48. IZMAT for symmetry internal coords
49. B matrix
50. B inverse matrix
51. overlap matrix in Lowdin basis,
temp Fock matrix storage for ROHF
52. genuine MOPAC overlap matrix
53. MOPAC repulsion integrals
54. Coulomb integrals
55-60. not used
61. temp MO storage for GVB and ROHF-MP2
62. temp density for GVB
63. dS/dx matrix for hessians
64. dS/dy matrix for hessians
65. dS/dz matrix for hessians
66. derivative hamiltonian for OS-TCSCF hessians
67. partially formed EG and EH for hessians
68. MCSCF first order density in MO basis
69. alpha Lowdin populations
70. beta Lowdin populations
71. alpha orbitals during localization
72. beta orbitals during localization
73-83. not used
84. d/dx dipole velocity integrals
85. d/dy dipole velocity integrals
86. d/dz dipole velocity integrals
87-88. reserved for effective fragment use
89. not used
90. ECP coefficients
int 91. ECP labels
92. ECP coefficients
int 93. ECP labels
94. bare nucleus Hamiltonian during FFIELD runs
95. x dipole integrals
96. y dipole integrals
97. z dipole integrals
98. former coords for Schlegel geometry search
99. former gradients for Schlegel geometry search
In order to correctly pass data between different
machine types when running in parallel, it is required
that a DAF record must contain only floating point values,
or only integer values. No logical or Hollerith data may
be stored. The final calling argument to DAWRIT and
DAREAD must be 0 or 1 to indicate floating point or
integer values are involved. The records containing
integers are so marked in the above list.
Physical record 1 (containing the DAF directory) is
written whenever a new record is added to the file. This
is invisible to the programmer. The numbers shown above
are "logical record numbers", and are the only thing that
the programmer need be concerned with.
