Changes between Version 8 and Version 9 of experiments


Ignore:
Timestamp:
10/03/08 13:26:41 (10 years ago)
Author:
daley
Comment:

tweaks

Legend:

Unmodified
Added
Removed
Modified
  • experiments

    v8 v9  
    3232 
    3333For the weak scaling we attempt to keep the number of blocks per processor  
    34 approximately constant. 
    35  
    36 The way to experiment is to use the setup as is, and change the 
    37 lrefine_max in flash.par to get to the next problem size. Keep adjusting the number of 
    38 processors until you get roughly the same number of blocks per processor. You may have less of a problem 
    39 doing this on XT4 since there is lots more memory per processor. Also, eliminate IO 
     34approximately constant.  In simpler problems, it is possible to just change the 
     35lrefine_max in flash.par to get to the next problem size.  Then keep adjusting the number of 
     36processors until you get roughly the same number of blocks per processor.  
     37However, with WD_Def the amount of refinement is hard to calculate when varying  
     38lrefine_max.  Therefore, we vary r_match & uni_radius in the flash.pars 
     39to adjust the problem size.  We do not vary the refinement maximum level. 
     40It required quite a bit of work to come up with these values. 
     41They were chosen to give problem sizes of (roughly) 100 * 64 * 2^n 
     42total blocks, for integer n, n is the number in scaling<n>. 
     43 
     44Note, weak scaling should be easier on the  
     45XT4 since there is lots more memory per processor. Also, eliminate IO 
    4046from the runs, that should give you more memory to play around in. 
    41  
    42 However, it seems for WD_def we do not vary lrefine_max.   
    43 Chris needs to speak to his work colleague to understand this  
    44 more thoroughly.  For the time being here is the email  
    45 describing the weak scaling for WD_Def: 
    46 {{{ 
    47 On Wed, 1 Oct 2008, Chris Daley wrote: 
    48  
    49 > Also can you please explain why we change: 
    50 > r_match 
    51 > refine_uni_radius 
    52 > in each flash.par? 
    53  
    54 THAT is how the size of the problem (i.e., muber of blocks) was 
    55 changed.  Note the lrefine_min and lrefine_max do not change! 
    56  
    57 > How do you know the correct value? 
    58  
    59 It required quite a bit of work to come up with these values. 
    60  
    61 They were chosen to give problem sizes of (roughly) 100 * 64 * 2^n 
    62 total blocks, for integer n, n is the number in scaling<n> I believe. 
    63  
    64 Klaus  
    65 }}} 
    66  
    6747 
    6848More details, which I include in their raw format: