Summary: CPU-MEM-RWF relation from G94



 > To check the effect of %MEM on G94 jobs, i did 4 CCSD(T) freq
 > calculations using same input file but different %MEM=32, 64, 128 and
 > 150MB. No other application prog. was running during these four jobs.
 >
 > The result is:
 >   %MEM      CPU       RWF(MB)       MaxMem (from output)
 >    32mb    55m 35.2s   45           4194304
 >    64mb    57m 11.8s   77           8388608
 >   128mb    59m 37.5s  141           16777216
 >   150mb   1h 0m 25.5s  163          19660800
 >
 > It looks odd. Expert's comment please.
 > Tapas
 > PS: I am using RS/6000 Model 590 and AIX4.2
 >     Total 256mb memory and 13(2+2+9)gb hard disk.
 >
 ****************************************************************************
 AIX uses the spare RAM memory as the disk cache.  Thus when you allocate
 memory to G94 with the %MEM directive you are also actually reducing the
 disk cache, slowing iour I/O. If your job is disk-intensive ( like CC ) it
 is reasonably possible to observe behaviors like the one you experience
 just my 0.02$
 Ivan
 --
 Dr. Ivan Rossi - Computational Chemistry Consultant
 e-mail: ivan : at : lipid.biocomp.unibo.it    phone: (+39)-051-456177
 ******************************************************************************
 Though not an expert, I may have some relevant observations to share.
 You point out something odd about the Gaussian programs, namely the
 less-than-intuitive relationship between the amount of memory allocated
 and
 the ressources used.  We have recently run a series of tests on a variety
 of machines to examine these relationships, but I'm afraid that in the end
 we were only confused, abeit on a higher level.
 We concentrated our efforts on time-consuming jobs, which in our line of
 work means MP2/4 and QCISD(T) calculations (G2 and CBS-Q methods).  The
 results are somewhat dependent upon the size of the system and upon the
 hardware employed.
         For a relatively small open-shell system (a C3H7N radical cation)
 we found that the UMP2/6-311+G(3df,2p) step of a G2(MP2) calculation on a
 single node of an IBM SP2 cluster would use within reason the same amount
 of cpu time, regardless of whether the job was allowed to use (%Mem) 4MW,
 16MW, or 64MW of memory.  The reported maximum size of the read-write file
 in the three runs was 700MB, 365MB, and 633MB; in two runs using 10MW and
 100MW we observed that the read-write file was almost three times bigger
 in
 the 100MW run than in the 10MW run (900MB vs 350MB).
         The QCISD(T)/6-311G(d,p) step ran 30% faster when 64MW were
 available (for the C3H7N+. system, the memory required for AO integral
 storage is 25MW); the time used for the triples was the same, regardless
 of
 the %Mem setting. (see below).
 Doing CBS-Q calculations on an SGI Origin200 for a C3H7O cation, we found
 that the execution times would vary only little with the amount of memory
 allocated (8MW, 16MW, 32MW, 48MW).  The memory required for in-memory
 allocated (1.33GB, 1.39GB, 1.51GB, 1.64GB).
 QCISD(T)/6-31G(d) calculations (Origin200) for a somewhat bigger system (a
 C4H12N cation) indicated that the QCISD step would benefit from increased
 memory (test runs with 8MW, 16MW, 32MW, 48MW; in-memory storage of AO
 integrals requiring 25MW), but that, strangely enough, the triples
 calculations would actually slow down.
 To examine this further we did QCISD(T)/6-31G(d) calculations for C5H14N
 cations on an SGI Origin200, on a single node of an IBM SP2 cluster, and
 on
 a Fujitsu VX1.  In all three instances we observed that the cpu-time for
 the triples calculations would INCREASE quite a bit with increased memory
 (data points: SGI 246 min (10MW), 321 min (60MW); IBM 97 min (16MW), 209
 min (80MW); the memory required for AO integral storage was 51MW; similar
 results on the Fujitsu machine).  These slow-downs were on all three
 machines to some extent off-set by a more rapid execution of the QCISD
 step
 with increased memory.  Also here we observed that increased memory would
 be accompanied by an increase in the size of the read-write file.
 Since doing these calculations I have been told that also other people
 have
 observed that the cpu-time used by the Gaussian programs for highly
 correlated calculations can show an inverse dependence on the amount of
 memory allocated.  I don't know if these things have changed in
 Gaussian98.
 Hope this helps.
 Steen
 --
 Steen Hammerum                                           steen : at : kiku.dk
 Department of Chemistry                              (+45) 35 32 02 08
 University of Copenhagen, Denmark               fax: (+45) 35 32 02 12
 ****************************************************************************
 --------------------------------------------
 Tapas Kar, Ph. D
 Asst. Scientist/Asst. Professor (Adjunct)
 Forestry Bldg 118
 Department of Chemistry
 Southern Illinois University at Carbondale
 Illinois 62901-4409
 Fax: (618) 453 6408
 Tel: (618) 453 6433(Lab) 6485(Office)
 --------------------------------------------