Skip to content
Prev 23293 / 29559 Next

GStat 'predict.c' parallel estimation

On 27/08/15 16:55, Edzer Pebesma wrote:
Thanks again Edzer. That's good to know about the global variables. Re 
the threaded C code working with R, my understanding was that the while 
loop undertaking the estimation is not interacting with R and so can be 
made parallel without (hopefully) influencing R. Does this sound reasonable?
To clarify, the best approach I see for simulations is to randomly 
select, say, 10*"number of cores"  grid cells and undertake their 
parallel estimation. These solutions are then added back into the "msim" 
and another 10*"number of cores" grid cells are then selected and 
estimated. This would overcome the boundary effects but would mean that 
the cores may not be running at 100% all of the  time (ie all 10*"number 
of cores" cells would need to finish before the next batch could start)
Thanks. I didn't know about circulant embedding. However, my problem is 
high non-bi-Gaussian. The extreme values within the system (which is 
water table elevation) are significantly more correlated than 'average' 
values.