(6 May 2010)
**********************************
* *
* Section 6 - Hardware Specifics *
* *
**********************************
This section of the manual contains pages dealing in a
general way with dynamic memory allocation in GAMESS, the
BLAS routines, and vectorization.
The remaining portions of this section consist of
specific suggestions for each type of machine. You should
certainly read the section pertaining to your computer. It
is a good idea to look at the rest of the machines as well,
as you may get some ideas! The directions for executing
GAMESS are given, along with hints and other tidbits. Any
known problems with certain compiler versions are described
in the control language files themselves, not here.
The currently supported machines are all running Unix.
The embedded versions for IBM mainframes and VAX/VMS have
not been used in many years, and are no longer described
here. There are binary versions for Windows available on
our web site, but we do not supply a source code version
for Windows (except that the Unix code will compile under
the Cygwin Unix environment for Windows). Please note that
with the OS X system, the Macintosh is considered to be a
system running Unix, and is therefore well supported.
Contents of this chapter:
Dynamic memory in GAMESS
BLAS routines
Vectorization of GAMESS
Notes for specific machines
------------------------------------------------------------
Dynamic memory in GAMESS
GAMESS allocates its working memory from one large pool
of memory. This pool consists of a single large array,
which is partitioned into smaller arrays as GAMESS needs
storage. When GAMESS is done with a piece of memory, that
memory is freed for other uses.
The units for memory are words, a term which GAMESS
defines as the length used for floating point numbers, 64
bits, that is 8 bytes per word.
GAMESS contains two memory allocation schemes. For
some systems, a primitive implementation allocates a large
array of a *FIXED SIZE* in a common named /FMCOM/. This is
termed the "static" implementation, and the parameter
MWORDS in $SYSTEM cannot request an amount larger than
chosen at compile time. Wherever possible, a "dynamic"
allocation of the memory is done, so that MWORDS can (in
principle) request any amount. The memory management
routines take care of the necessary details to fool the
rest of the program into thinking the large memory pool
exists in common /FMCOM/.
Computer systems which have "static" memory allocation
are IBM mainframes running VM or MVS to which we have no
direct access for testing purposes. If your job requires a
larger amount of memory than is available, your only
recourse is to recompile UNPORT.SRC after choosing a larger
value for MEMSIZ in SETFM.
Computer which have "dynamic" memory allocation are all
Unix systems and VMS. In principle, MWORDS can request any
amount you want to use, without recompiling. In practice,
your operating system will impose some limitation. As
outlined below, common sense imposes a lower limit than
your operating system will.
By default, most systems allocate a small amount of
memory: one million words. This amount is quite small by
modern standards, and therefore exists on all machines. It
is left up to you to increase this with your MWORDS input
to what your machine has. EXETYP=CHECK runs will always
tell you the amount of memory you need.
Many computations in GAMESS implement out of memory
algorithms, whenever the in memory algorithm can require an
excessive amount. The in memory algorithms will perform
very poorly when the work arrays reside in virtual memory
rather than physical memory. This excessive page faulting
activity can be avoided by letting GAMESS choose its out of
core algorithms. These are programmed such that large
amounts of numbers are transferred to and from disk at the
same time, as opposed to page faulting for just a few
values in that page. So, pick an amount for MWORDS that
will reside in the physical memory of your system! MWORDS,
multiplied by 8, is roughly the number of Mbytes and should
not exceed more than about 90% of your installed memory
(less if you are sharing the computer with other jobs!).
The routines involved in memory allocation are VALFM,
to determine the amount currently in use, GETFM to grab a
block of memory, and RETFM to return it. Note that calls
to RETFM must be in exactly inverse order of the calls to
GETFM. SETFM is called once at the beginning of GAMESS to
initialize, and BIGFM at the end prints a "high water mark"
showing the maximum memory demand. GOTFM tells how much
memory is not yet allocated.
BLAS routines
The BLAS routines (Basic Linear Algebra Subprograms)
are designed to perform primitive vector operations, such
as dot products, or vector scaling. They are often found
implemented in a system library, even on scalar machines.
If this is the case, you should use the vendor's version!
The BLAS are a simple way to achieve BOTH moderate
vectorization AND portability. The BLAS are easy to
implement in FORTRAN, and are provided in the file BLAS.SRC
in case your computer does not have these routines in a
library.
The BLAS are defined in single and double precision,
e.g. SDOT and DDOT. The very wonderful implementation of
generic functions in FORTRAN 77 has not yet been extended
to the BLAS. Accordingly, all BLAS calls in GAMESS use the
double precision form, e.g. DDOT. The source code
activator translates these double precision names to single
precision, for machines such as Cray which run in single
precision.
If you have a specialized BLAS library on your machine,
for example IBM's ESSL, Compaq's CXML, or Sun's Performance
Library, using them can produce significant speedups in
correlated calculations. The compiling scripts attempt to
detect your library, but if they fail to do so, it is easy
to use one:
a) remove the compilation of 'blas' from 'compall',
b) if the library includes level 3 BLAS, set the value
of 'BLAS3' to true in 'comp',
c) in 'lked', set the value of BLAS to a blank, and
set libraries appropriately, e.g. to '-lessl'.
Check the compilation log for mthlib.src, in particular, to
be sure that your library is being found. It has a
profound effect on the speed of MP2 and CC computations!
The reference for the level 1 BLAS is
C.L.Lawson, R.J.Hanson, D.R.Kincaid, F.T.Krogh
ACM Trans. on Math. Software 5, 308-323(1979)
Vectorization of GAMESS
As a result of a Joint Study Agreement between IBM and
NDSU, GAMESS has been tuned for the IBM 3090 vector
facility (VF), together with its high performance vector
library known as the ESSL. This vectorization work took
place from March to September of 1988, and resulted in
a program which is significantly faster in scalar mode, as
well as one which can take advantage (at least to some
extent) of a vector processor's capabilities. Since our
move to ISU we no longer have access to IBM mainframes,
but support for the VF, as well as MVS and VM remains
embedded within GAMESS. Several other types of vector
computers are supported as well.
Anyone who is using a current version of the program,
even on scalar machines, owes IBM their thanks both for
NDSU's having had access to a VF, and the programming time
to do code improvements in the second phase of the JSA,
from late 1988 to the end of 1990.
Some of the vectorization consisted of rewriting loops
in the most time consuming routines, so that a vectorizing
compiler could perform automatic vectorization on these
loops. This was done without directives, and so any
vectorizing compiler should be able to recognize the same
loops.
In cases where your compiler allows you to separate
scalar optimization from vectorization, you should choose
not to vectorize the following sections: INT2A, GRD2A,
GRD2B, and GUGEM. These sections have many very small
loops, that will run faster in scalar mode. The remaining
files will benefit, or at least not suffer from automatic
compiler vectorization.
The highest level of performance, obtained by
vectorization at the matrix level (as opposed to the
vector level operations represented by the BLAS) is
contained in the file VECTOR.SRC. This file contains
replacements for the scalar versions of routines by the
same names that are contained in the other source code
modules. VECTOR should be loaded after the object code
from GAMESS.SRC, but before the object code in all the
other files, so that the vector versions from VECTOR are
the ones used.
Most of the routines in VECTOR consist of calls to
vendor specific libraries for very fast matrix operations,
such as IBM's Engineering and Scientific Subroutine
Library (ESSL). Look at the top of VECTOR.SRC to see
what vector computers are supported currently.
If you are trying to bring GAMESS up on some other
vector machine, do not start with VECTOR. The remaining
files (excepting BLAS, which are probably in a system
library) represent a complete, working version of GAMESS.
Once you have verified that all the regular code is
running correctly, then you can adapt VECTOR to your
machine for the maximum possible performance.
Vector mode SCF runs in GAMESS on the IBM 3090 will
proceed at about 90 percent of the scalar speed on these
machines. Runs which compute an energy gradient may
proceed slightly faster than this. MCSCF and CI runs
which are dominated by the integral transformation step
will run much better in vector mode, as the transformation
step itself will run in about 1/4 time the scalar time on
the IBM 3090 (this is near the theoretical capability of
the 3090's VF). However, this is not the only time
consuming step in an MCSCF run, so a more realistic
expectation is for MCSCF runs to proceed at 0.3-0.6 times
the scalar run. If very large CSF expansions are used
(say 20,000 on up), however, the main bottleneck is the CI
diagonalization and there will be negligible speedup in
vector mode. Several stages in an analytic hessian
calculation benefit significantly from vector processing.
A more quantitative assessment of this can be reached
from the following CPU times obtained on a IBM 3090-200E,
with and without use of its vector facility:
ROHF grad RHF E RHF hess MCSCF E
------- ------ ------- ------
scalar 168 ( 1 ) 164 ( 1 ) 917 ( 1 ) 903 ( 1 )
vector 146 (0.87) 143 (0.87) 513 (0.56) 517 (0.57)
Notes for specific machines
GAMESS will run on many kinds of UNIX computers. These
systems runs the gamut from very BSD-like systems to very
ATT-like systems, and even AIX. Our experience has been
that all of these UNIX systems differ from each other. So,
putting aside all the hype about "open systems", we divide
the Unix world into four classes:
Supported: Apple MAC under OS X, HP/Compaq/DEC AXP, HP
PA-RISC, IBM RS/6000, 64 bit Intel/AMD chips such as the
Xeon/Opteron/Itanium, and Sun ultraSPARC. These are the
only types of computer we currently have at ISU, so these
are the only systems we can be reasonably sure will work
(at least on the hardware model and O/S release we are
using). Both the source code and control language is
correct for these.
Acquainted: Cray XT, IBM SP, SGI Altix/ICE, and SGI
MIPS. We don't have any of these systems at ISU, so we
can't guarantee that these work. GAMESS has been run on
each of these offsite, perhaps recently, but perhaps not.
The source code for these systems is probably correct, but
the control language may not be. Be sure to run all the
test cases to verify that the current GAMESS still works on
these brands.
Jettisoned: Alliant, Apollo, Ardent, Celerity, Convex,
Cray T3E, Cray vectors, DECstations, FPS model 500, Fujitsu
AP and VPP, HP Exemplar, Hitachi SR, IBM AIX mainframes,
Intel Paragon, Kendall Square, MIPS, NCube, and Thinking
Machines. In most cases the company is out of business, or
the number of machines in use has dropped to near zero. Of
these, only the Celerity version's death should be mourned,
as this was the original UNIX port of GAMESS, back in July
1986.
Terra Incognita: everything else! You will have to
decide on the contents of UNPORT, write the scripts, and
generally use your head.
* * * * *
You should have a file called "readme.unix" at hand
before you start to compile GAMESS. These directions
should be followed carefully. Before you start, read the
notes on your system below, and read the compiler clause
for your system in 'comp', as notes about problems with
certain compiler versions are kept there.
Execution is by means of the 'rungms' script, and you
can read a great deal more about its DDIKICK command in the
installation guide 'readme.ddi'. Note in particular that
execution of GAMESS now uses System V shared memory on many
systems, and this will often require reconfiguring the
system's limits on shared memory and semaphores, along with
a reboot. Full details of this are in 'readme.ddi.
Users may find examples of the scalability of parallel
runs in the Programmer's Reference chapter of this manual.
* * * * * *
AMD Opteron and other chips: see "linux64" below.
AXP: These are scalar systems. This category means
any AXP machines, whether labeled Digital or Compaq or HP
on the front, with an O/S called OSF1, Digital Unix, or
Tru64. It also includes systems running Linux, see below.
The unique identifier is therefore the AXP chip, so the
compiling target is 'axp64', rather than a company name.
The compiling script invokes the f77 compiler, so read
'comp' if you have the f90 compiler instead. This version
was changed to use native 64 bit integers in fall 1998.
You can also run GAMESS on AXP Linux, by using the
Tru64 Compaq compilers, which permit the Tru64 version to
run. Do not use g77 which allocates 32 bit integers, as
the system's malloc routine for dynamic memory allocation
returns 64 bit addresses, which simply cannot be stored in
32 bit integers. The Compaq compilers can easily generate
64 bit integers, so obtain FORTRAN and C from
http://h18000.www1.hp.com/products/software/alpha-tools
Then compile and link using target 'compaq-axp'.
Cray XT: a massively parallel platform, based on dual
Opteron processor blades connected by Cray's 3D mesh,
running a node O/S called "Compute Node Linux". The
message passing involves a DDI running over MPI with a user
selectable number of data servers. Unfortunately, the DDI
is not fully integrated into our main code yet, and the
scripting is a bit rough. Good support for these (XT3
through XT6) is expected by summer 2010.
Digital: See AXP above.
HP: Any Intel or PA-RISC series workstation or server.
Help with this version has come from due to Fred Senese,
Don Phillips, Tsuneo Hirano, and Zygmunt Krawczyk. Dave
Mullally at HP has been involved in siting HP systems at
ISU, presently Itanium2 based. So, we used 'hpux32' for
many years, but are now running only the 'hpux64' version.
The latter version can be considered to be carefully
checked since it is in use at ISU, but please be a little
more careful checking tests if you use 'hpux32'.
IBM: "superscalar" RS/6000. There are two targets for
IBM workstations, namely "ibm32" and "ibm64", neither of
these should be used on a SP system. Parallelization is
achieved using the TCP/IP socket calls found in AIX.
IBM-SP: The SP parallel systems. This is a 64 bit
implementation. The new DDI library will operate with LAPI
support for one-sided messaging, and a special execution
script for LoadLeveler is included.
IBM Blue Gene: This target is "ibm-bg". The older BG/L
has been outmoded by the BG/P, but we still have an "L" at
ISU. These are massively parallel machine, using a 32 bit
PowerPC, and a limited amount of node memory. The "L" uses
DDI running over the ARMCI library, running in turn over
MPI, so the "L" does not use data servers. The "P" uses a
straightforward DDI to MPI interface, with data servers.
The "L" port was done by Brian Smith of IBM and Brett Bode
at ISU, included in GAMESS in June 2005, and changed to use
ARMCI in 2007 by Andrey Asadchev of ISU. Nick Nystrom's
initial port to the "P" system was polished up by Graham
Fletcher at Argonne National Labs in 2010. Special notes,
and various files to be used on this system are stored in
the directory ~/gamess/machines/ibm-bg.
Linux32: this means any kind of 32 bit chips, but
typically is used only when "uname -p" replies "x86".
Nearly every other chip is 64 bits, so see also Linux64
just below. This version is originally due to Pedro
Vazquez in Brazil in 1993, and modified by Klaus-Peter
Gulden in Germany. The usefulness of this version has
matched the steady growth of interest in PC Unix, due to
the improvement in CPU, memory, and disks, to workstation
levels. We acquired a 266 MHz Pentium-II PC running RedHat
Linux in August 1997, and found it performed flawlessly.
In 1998 we obtained six 400 MHz Pentium-IIs for sequential
use, and in 1999 a 16 PC cluster, running in parallel day
in and day out. We have used RedHat 4.2, 5.1, 6.1, 7.1,
and Fedora Core 1, prior to switching over exclusively to
64-bit Linux. This version is based on gfortran or g77,
gcc, and the gcclib, so it should work for any kind of 32
bit Linux. This version uses 'sockets' for its message
passing. The configuration script will suggest possible
math library choices to you.
By 2010, probably most Linux systems in existence are
64-bit capable, so the next version is more better!
Linux64: this means any sort of 64 bit chip running an
appropriate 64 bit Linux operating system. The most common
"linux64" build is on AMD or Intel chips, where "uname -p"
returns x86_64 or ia64. However, if you choose the
'gfortran' compiler, no processor-specific compiler flags
are chosen, so this version should run on any 64-bit Linux
system, e.g. AXP or SPARC.
If you are running on Intel/AMD processors, the
configuration script lets you choose various FORTRAN
compilers: GNU's gfortran, Intel's ifort, Portland Group's
pgf77, and Pathscale's pathf90. You can choose a variety
of math libraries, such as Intel's MKL, AMD's ACML, or
ATLAS. You can choose to use MPI if your machine has a
good network for parallel computing, with the options for
the MPI type specified in detail in the file, but sockets
are an easy to use alternative to MPI.
The choices for FORTRAN, math library, and MPI library
can all be "mixed and matched". Except for 'gfortran',
almost all this software has to be added to a standard
Linux distribution. It is your responsibility to install
what you want to use, to set up execution paths, to set up
run time library paths (LD_LIBRARY_PATH), and so forth.
The 'config' script will need to ask where these software
packages are installed, since your system manager may have
placed them almost anywhere.
Macintosh OS X: This is for Apple running OS X, which
is a genuine Unix system "under the hood". This version
closely resembles the Linux version. Installation of
Apple's XCODE (from the OS X distribution DVD) gives you a
C compiler and a math library. You can obtain a FORTRAN
compiler (gfortran for 64 bit or g77 for 32 bits) from the
wonderful web site of Gourav Khanna:
http://hpc.sourceforge.net
Request target "mac32" if your OS X is 10.4, or "mac64" if
your OS X is 10.5 or newer.
NEC SX: vector system. This port was done by Janet
Fredin at the NEC Systems Laboratory in Texas in 1993, and
she periodically updates this version, including parallel
usage, most recently in Oct. 2003. You should select both
*UNX and *SNG when manually activating ACTVTE.CODE, and
compile actvte by "f90 -ew -o actvte.x actvte.f".
Silicon Graphics: The modern product line of this
company is called Altix or Altix ICE. The operating system
is Linux, the chips are 64 bit Intel processors, and the
natural compiler and math library choices are Intel's ifort
and MKL. The SGI software ProPack turns these commodity
components into a supercomputer, and DDI should use the MPI
library 'mpt' found in ProPack. Accordingly, the compiling
target should be 'linux64', selecting ifort, MKL, and then
mpt.
Silicon Graphics: The ancient product line of this
company use various MIPS chips such as R4x00, R5000,
R12000, etc. There are very few of these machines left, so
target's 'sgi32' and 'sgi64' should be regarded as "rusty".
The 32 bit target uses sockets communications, while the 64
bit one will use an old SHMEM interface, that only
partially implements DDI. Hence FMO and parallel CCSD(T)
will not run on 'sgi64'.
Sun: scalar system. This version is set up for the
ultraSPARC or Opteron chips, running Solaris. The target
for either chip is "sun64" as the scripts can automatically
detect which one you are using, and adjust for that. Since
Sun provided a ultraSPARC E450 system in 1998, two
ultraSPARC3 Sunfire 280R systems in 2002, and a Opteron
V40Z system in 2006, to the group at Iowa State, the Sun
version is very reliable. Install the SunPerf math library
from the compiler suite for maximum BLAS performance.
Parallelization is accomplished using TCP/IP sockets and
SystemV shared memory.
Edited by Shiro KOSEKI.