Fitting becoming very slow for ternary systems

[YuanbinLiu]

Hi all,

When using ACE to fit ternary systems, I‘ve noticed that the speed becomes very slow. Here are some parameters I used; could it be because the parameters are set too high or some reasons else?

model = acemodel(elements=[:Ba, :Zr, :S], 
order=3, 
totaldegree=16, 
rcut=5.0,
Eref=[:Ba => -0.03428427, :Zr => -1.34987151, :S => -0.09468838],
)

It has been stuck in here for more than 2 hours:

[ Info:   - Assembly completed.
[ Info: Assembling full weight vector.

Progress:   0%|                                         |  ETA: 0:02:07
Progress:   8%|███▏                                     |  ETA: 0:00:04
Progress:  57%|███████████████████████▍                 |  ETA: 0:00:00
Progress: 100%|█████████████████████████████████████████| Time: 0:00:00

[cortner]

this is very very strange. And certainly not a problem with the ternary systems. Do you mind posting the size of the design matrix?

[cortner]

Any way to do the SVD distributed?

[YuanbinLiu]

Thanks for the suggesitons! I will have a try soon.

[YuanbinLiu]

Could 128 processors significantly imporve computaional efficiency over 32 processors? What is the recommended numer of processors?

[cortner]

Shared memory? That should help a lot. But it's still a cubic scaling algorithm. Try LSQR and also RRQR.

And if RRQR works for you then we should make a sketched SVD an option for BLR

[YuanbinLiu]

Ok, will let you know if it works