speed of new cp2k

[Matt W]

Hi Limin,

I'm a little surprised it's slower using that cpu / processor ratio.
Are you running bulk, or do you have a surface in the system?  If you
were running with PRINT_LEVEL MEDIUM or higher could you post the
distribution of the grids, i.e. something like for any distributed
grids

 RS_GRID: Information for grid
number                                          1
 RS_GRID:   Bounds   1            -54      53
Points:         108
 RS_GRID:   Bounds   2            -54      53
Points:         108
 RS_GRID:   Bounds   3           -108     107
Points:         216
 RS_GRID: Real space distribution
over                                  8 groups
 RS_GRID: Real space distribution along
direction                              1
 RS_GRID: Border
size                                                         17
 RS_GRID:   Distribution                         Average
Max         Min
 RS_GRID:   Planes                                  47.5
48          47

There have also been some additional changes to the code (the last
about 15 minutes ago) that improve the performance of the new
routines(hoepfully).  I would suggest updating (once again) and see if
the performance is still as you describe.

Please let us know how it goes.

Cheers,

Matt



On Jan 17, 3:47 pm, "lml... at gmail.com" <lml... at gmail.com> wrote:
> hi,
>
> It seems that the speed of cp2k becomes a little lowly (about 5%)
> instead of speeding up after Jan. 10 for the new 'integrate and
> collocate routines'.  Is this general? Do I need set new parameter for
> the 'distributed realspace grids' part compared with the old one?
> I test on the XT4 with PGI 7 using 32 water using 32 CPUs.
>
> Thanks.
> Limin