$SYSTEM group         (optional)
 
    This group provides global control information for
your computer's operation.  This is system related input,
and will not seem particularly chemical to you!
 
MWORDS =  the maximum replicated memory which your job can
          use, on every core.  This is given in units of
          1,000,000 words (as opposed to 1024*1024 words),
          where a word is defined as 64 bits.  (default=1)
 
In case finer control over the replicated memory is needed,
this value can be given in units of words, with the old
keyword MEMORY, instead of MWORDS.
 
MEMDDI =  the grand total memory needed for the distributed
          data interface (DDI) storage, given in units of
          1,000,000 words. See Chapter 5 of this manual for
          an extended explanation of running with MEMDDI.
 
note: the memory required on each processor core for a run
      using p cores is therefore MEMDDI/p + MWORDS.
 
The parallel runs that currently require MEMDDI are:
      SCFTYP=RHF   MPLEVL=2 energy or gradient
      SCFTYP=UHF   MPLEVL=2 energy or gradient
      SCFTYP=ROHF  MPLEVL=2 OSPT=ZAPT energy or gradient
      SCFTYP=MCSCF MPLEVL=2 energy
      SCFTYP=MCSCF using the FULLNR or JACOBI convergers
      SCFTYP=MCSCF analytic hessian
      SCFTYP=any   CITYP=ALDET, ORMAS, GUGA
      SCFTYP=any   energy localization
      SCFTYP=RHF   CCTYP=CCSD or CCSD(T)
All other parallel runs should enter MEMDDI=0, for they use
only replicated memory.
Some serial runs execute the parallel code (on just 1 CPU),
for there is only a parallel code.  These serial runs must
give MEMDDI as a result:
      SCFTYP=ROHF  MPLEVL=2 OSPT=ZAPT gradient/property run
      SCFTYP=MCSCF analytic hessian
Two kinds of runs (RI-MP2 and parallel CCSD(T)) use an
additional type of memory, for which there is no input
keyword.  Please read EXETYP=CHECK output carefully to
learn the total memory/node requirements for these two!
 
TIMLIM =  time limit, in minutes.  Set to about 95 percent
          of the time limit given to the batch job (if you
          use a queueing system) so that GAMESS can stop
          itself gently.  (default=525600.0 minutes)
 
PARALL =  a flag to cause the program to execute the
          parallel algorithm, in cases where different
          serial and parallel codes exist, if you happen to
          be running on only one core.
          The default is .FALSE. if you are running on one
          core. The main purpose of this keyword is to
          allow you to do EXETYP=CHECK runs on only one
          core, when your intent is perform the actual
          calculation in parallel.
          PARALL is ignored for runs on more than one core,
          when of course parallel algorithms are executed.
 
KDIAG  =    diagonalization control switch
       = 0  use a vectorized diagonalization routine
            if one is available on your machine,
            else use EVVRSP. (default)
       = 1  use EVVRSP diagonalization.  This may
            be more accurate than KDIAG=0.
       = 2  use GIVEIS diagonalization
            (not as fast or reliable as EVVRSP)
       = 3  use JACOBI diagonalization
            (this is the slowest method)
       = 4  use DSYEV diagonalization from LAPACK
       = 5  use DSYEVD for DFTB Fock matrices and
            DSYEV elsewhere (both from LAPACK).
 
            To use LAPACK, you should link it.
            You may be able to edit comp and lked to
            take advantage of your LAPACK library.
            Some math libraries such as MKL, OpenBLAS,
            and Fujitsu SSL come with LAPACK.
 
COREFL =  a flag to indicate whether or not GAMESS
          should produce a "core" file for debugging
          when subroutine ABRT is called to kill
          a job.  This variable pertains only to
          UNIX operating systems.  (default=.FALSE.)
 
BALTYP = Parallel load balance scheme:
       = SLB uses  static load balancing.
       = DLB uses dynamic load balancing (default).
         Dynamic load balancing attempts to spread out
         possibly unequal work assignments based on the
         rate at which different nodes complete tasks.
               For historical reasons, it is permissible
               to spell SLB as LOOP, and DLB as NXTVAL.
 
MXSEQ2 = 300 (default)
MXSEQ3 = 150 (default)
         Matrix/vector problem size in loops requiring
         either O(N**2) or O(N**3) work, respectively.
         Problems below these sizes are run purely serial,
         to avoid poor communication/computation ratios.
 
NODEXT = array specifying node extensions in GDDI for each
         file.  Non-zero values force no extension.
         E.g., NODEXT(40)=1 forces file 40 (file numbers
         are unit numbers used in GAMESS, see "rungms" or
         PROG.DOC) to have the name of $JOB.F40 on all
         nodes, rather than $JOB.F40, $JOB.F40.001,
         $JOB.F40.002 etc.
         This option should not be used on multicore
         machines, and is really outdated now.
         (default: all zeros)
 
IOSMP  = Parallelise I/O on SMP machines with multiple hard
         disks.  Two parameters are specified, whose
         meaning should be clear from the example.
              iosmp(1)=2,6
         2 refers to the number of HDDs per SMP box.
         6 is the location of the character in the file
         names that switches HDDs, i.e. if HDDs are mounted
         as /work1 and /work2, then 6 refers to the
         position of the number 1 in /work1. The file
         system should permit disks attached with directory
         names differing by one symbol.
         (default: 0,0, disable the feature)
 
MODIO  = Global I/O options (bitwise additive)
                (default: 0)
         1 - forbid flushing files
         2 - do not close dictionary file in GDDI
             (only record indices are reset)
         4 - do not print timings on each rank at run end.
         8 - forbid grid data saving in DFT
             (prevent F22 from being opened;
             however, in CPKS F22 is always opened).
        16 - reduce I/O (skip various printing)
        32 - do not open file F15 (in SOSCF) on slaves
             and do not close on all.
        64 - use in-memory F15 (in SOSCF). This also
             parallelizes one more step in SOSCF.
       128 - always run EVVRSP sequentially. This is useful
             on mixed CPU type clusters.
       256 - reduce timing output.
       512 - use XYZ file to store coordinates.
      1024 - enforce minimal output
      2048 - a single flag to turn on various accceleration
             options
      4096 - redirect OUTPUT and PUNCH output to /dev/null
             on all parallel ranks except the global master.
             Use only if OS permits it (a UNIX).
             Mainly intended for GDDI runs to prevent masters
             from writing anything. It is assumed that the
             global master writes all important data.
 
MEM10  = words used to store dictionary file F10 in memory.
         Selecting this option will replace any I/O for F10
         by memory access.
         Default: 0 (disk-based F10)
 
MEM22  = words used to store grid file F22 for DFT in memory.
         Selecting this option will skip any I/O for F22.
         Mainly useful for DFT Hessian and analytic FMO-DFT
         gradient, both using CPKS. MEM22 is not used for
         F22 in DFT energy/gradient (use MODIO=8 for removing
         I/O to F22).
         Default: 0 (disk-based F22)
 
NALIGN = align dynamic memory blocks so that relative addresses
         are multiples of NALIGN (absolute addresses are
         given by OS, usually a multiple of 64 bits).
         Default: 1 (effectively, no allignment)