[CPMD-list] PIMD: EHAM depends on number of processors ?

[Łukasz Walewski]

Dear CPMD users,

When playing with path integral MD I observed strange behaviour: the 
EHAM value is conserved during the simulation (as it should be) however 
its mean value depends on the number of processors the job runs on. I 
run several tests with the following results:

version, arch, no. of proc., EHAM (a.u.)
3.9.1, Intel, NP=1: -49.9124
3.9.2, Intel, NP=1: -49.9124
3.9.2, Cray, NP=1: -49.9130
3.9.1, Cray, NP=4: -49.9080
3.9.2, Cray, NP=4: -49.9080
3.9.1, Opteron, NP=1: -49.9117
3.9.1, Opteron, NP=4: -49.9066
3.9.1, Opteron, NP=8: -49.9166

The largest difference occurs for 3.9.1 on the Opteron cluster between 4 
and 8 processors: 6.3 kcal/mol ! Version number does not not play any 
role (at least on serial PC and Cray).

Is it due to numerical errors (like order of summation and other known 
nondeterministic aspects of parallel computing) or do you suspect 
problems somewhere else ? Are the results feasible ? correct ? Are other 
properties of the system affected ?

It may be important that all above simulations were carried out from the 
same input file attached below (most of it was inspired by 
CPMD-test/path-int/inp-2). The wavefunction was initially optimised and 
the configuration of the centroids was generated using DEBROGLIE 
CENTROID scheme.

With warm regards,
Lukasz


&CPMD
   PATH INTEGRAL
   MOLECULAR DYNAMICS CP
   ODIIS
     5
   RESTART WAVEFUNCTION COORDINATES VELOCITIES LATEST
   ISOLATED MOLECULE
   EMASS
    350.0
   TEMPERATURE
     300.D0
   NOSE IONS MASSIVE
     300.0  2500.0
   NOSE ELECTRONS
     2.D-4 15000.D0
   TIMESTEP
     1.D0
   MAXSTEP
    13781
   TRAJECTORY XYZ SAMPLE
     13
&END

&SYSTEM
   ANGSTROM
   SYMMETRY
     1
   CELL
     10.0 1.0 1.0 0 0 0
   CUTOFF
     35.0
&END

&ATOMS

*pp/O_MT_PBE.psp  KLEINMAN-BYLANDER
    LMAX=P
2
     1.184893   -0.892757    0.000000
    -1.184893   -0.892757    0.000000

*pp/C_MT_PBE.psp  KLEINMAN-BYLANDER
    LMAX=P
3
     1.187504    0.388446    0.000000
     0.000000    1.130651    0.000000
    -1.187504    0.388446    0.000000

*pp/H_MT_PBE.psp   KLEINMAN-BYLANDER
    LMAX=S
4
     0.000000   -1.136055    0.000000
     2.165818    0.881532    0.000000
     0.000000    2.211832    0.000000
    -2.165818    0.881532    0.000000

&END

&DFT
   NEWCODE
   FUNCTIONAL PBE
&END

&PIMD
   TROTTER DIMENSION
     8
   NORMAL MODES
     1.D0
   FACMASS
     1.D0
   DEBROGLIE CENTROID
     300.D0
   PRINT LEVEL
     1
&END

-- 
--------------------------------
Łukasz Walewski <ljw at icm.edu.pl>