obs+noise too big?

[wkitlasten]

pestpp-ies is "drawing observation noise realizations: 300" for 4.4 million obs. I have outsourced cognitive function to Claude, which claims it should take seconds to draw and be roughly 10GB. However, it is taking over 10 min and comes out at over 20GB. Is there something I can do to speed it up?

I have 100 cores and heaps of memory available.

The closest discussion I could find is: #312 (comment)

[jtwhite79]

Assuming its all uncorrelated, might try setting draw by groups to false....or let claude generate for you haha

[wkitlasten]

Yep, all uncorrelated. Turns out the big obs+noise wasn't the issue, it was just the last bit of information I could find. Here is a breakdown of what I can see as far as memory usage (mostly via watching task manager):

Ensembles:

obs+noise: 22GB
par ens: 2GB

pestpp-ies:

pestpp-ies (noptmax=0, sans ensembles): 6GB
pestpp-ies each agents (noptmax=1): 4.2GB

actual model:

forward_run.py (modflow api): 0.5MB

Peak usage via psutil running pestpp-ies with noptmax=1 using 3 agents and unavoidable corporate overhead (typically 10-15GB):

peak usage : 68GB

So the killer seems to be 300 agents * 4.2GB. Now I'm wondering if there is anything I can do about the 4.2GB of memory on each agent when the forward run only uses 0.5MB. Any clarification, suggestions, and/or workarounds would be appreciated.

[jtwhite79]

Ive tried to minimize the workers mem many times. The biggest consumer is the string names on the workers....are you using verbose obs and par names?

[wkitlasten]

That's useful, I will work with that. Cheers