edmf random seed not too random?

[mathomp4]

@narnold1 A question. While talking with NVIDIA about their GPU efforts with edmf we looked at this code:

GEOSgcm_GridComp/GEOSagcm_GridComp/GEOSphysics_GridComp/GEOSturbulence_GridComp/edmf.F90

Lines 347 to 356 in b3b07ba

	! get Poisson P(dz/L0)
	seedmf(1) = 1000000 * ( 100*thl(kte) - INT(100*thl(kte)))
	seedmf(2) = 1000000 * ( 100*thl(kte-1) - INT(100*thl(kte-1)))
	THE_SEED(1)=seedmf(1) + seedmf(2)
	THE_SEED(2)=seedmf(1) + seedmf(2)
	THE_SEED(1)=THE_SEED(1)*seedmf(1)/( seedmf(2) + 10)
	THE_SEED(2)=THE_SEED(2)*seedmf(1)/( seedmf(2) + 10)
	if(THE_SEED(1) == 0) THE_SEED(1) = 5
	if(THE_SEED(2) == 0) THE_SEED(2) = -5

and then the_seed is passed to Poisson:

GEOSgcm_GridComp/GEOSagcm_GridComp/GEOSphysics_GridComp/GEOSturbulence_GridComp/edmf.F90

Line 362 in b3b07ba

call Poisson(1,Nup,kts,kte,ENTf,ENTi,the_seed)

The thing is...I'm pretty sure the_seed(1) equals the_seed(2) unless they are 0. We were thinking maybe it should be:

 THE_SEED(1)=seedmf(1) + seedmf(2) 
 THE_SEED(2)=seedmf(1) - seedmf(2) 

at the beginning?

Just wondering.

Cc: @tclune as he probably knows more about algorithmically making random numbers than I do

[tclune]

It all looks very ad hoc to me. Some attempt to pre-randomize a seed?

[narnold1]

This section of code was copied from ras.F90. I'm not sure what motivated lines 351-356. Is THE_SEED=0 disallowed? I think the current code works, in the sense of producing randomness and satisfying regression tests, but I will double check that.

[mathomp4]

This section of code was copied from ras.F90. I'm not sure what motivated lines 351-356. Is THE_SEED=0 disallowed? I think the current code works, in the sense of producing randomness and satisfying regression tests, but I will double check that.

@narnold1 There are compilers where setting a seed that's not too random causes issues (see https://forums.developer.nvidia.com/t/random-number-seed-oddities/133398 by some "Matt Thompson"), so I always discourage that once I realized that. I mean, setting the seed based on input temperatures isn't bad, it's just like the seed construction was a bit oddly coded. 😄

[narnold1]

This issue is addressed by PR #783