Welcome, Guest

Author Topic: Kalassak runs plasim  (Read 877 times)

atomic7732

  • Global Moderator
  • *****
  • Posts: 3794
  • caught in the river turning blue
    • Paladin of Storms
Kalassak runs plasim
« on: April 28, 2019, 02:51:51 AM »
the long awaited sequel to WRF

idk have a monologue from irc i'm going to bed

these are more notes for myself:
[08:35]   iris-discord2   <latenivenatrix9953> oh there we go
[08:35]   iris-discord2   <latenivenatrix9953> that fixed it
[08:35]   iris-discord2   <latenivenatrix9953> conclusion:
[08:36]   iris-discord2   <latenivenatrix9953> mpich was either insufficient or somehow borked, so compiling with it did not allow plasim to recognize it was being run in parallel
[08:36]   iris-discord2   <latenivenatrix9953> even executing mpiexec -np 8 would initiate 8 independent threads that would complain that they were 1 thread each
[08:36]   iris-discord2   <latenivenatrix9953> that was causing the abort, here:
[08:37]   iris-discord2   <latenivenatrix9953> if (nproc .ne. NPRO .and. mypid == NROOT) then write(nud,)'Compiled for ',NPRO,' nodes' write(nud,)'Running on ',nproc,' nodes' call mpi_abort(myworld,mpinfo,mpinfo)
[08:38]   iris-discord2   <latenivenatrix9953> then
[08:38]   iris-discord2   <latenivenatrix9953> using mpimod_multi alleviated that problem, but caused the issue witht he weird independent namelists and the fact that the temperature data was only showing up in 1/8th of the entire grid
[08:38]   iris-discord2   <latenivenatrix9953> because it was running only 1/8th
[08:39]   iris-discord2   <latenivenatrix9953> so now that's fixed
[08:39]   iris-discord2   <latenivenatrix9953> by just editing mpimod.f90 to rename the mpi_info function to mpi_info2
[08:39]   iris-discord2   <latenivenatrix9953> and editing the corresponding function calls in icemod.f90 and oceanmod.f90

speed:
[08:42]   iris-discord2   <latenivenatrix9953> yikerino it's slow af
[08:44]   iris-discord2   <latenivenatrix9953> all my cpus are at 100%
[08:44]   iris-discord2   <latenivenatrix9953> we're at 2 hours
[08:45]   iris-discord2   <latenivenatrix9953> in what 10 mins
[08:47]   iris-discord2   <latenivenatrix9953> ooh
[08:47]   iris-discord2   <latenivenatrix9953> at that rate
[08:47]   iris-discord2   <latenivenatrix9953> a year will take 43200 mins
[08:47]   iris-discord2   <latenivenatrix9953> or 720 hours
[08:47]   iris-discord2   <latenivenatrix9953> or 30 days
[08:48]   iris-discord2   <latenivenatrix9953> hm
[08:49]   iris-discord2   <latenivenatrix9953> so theoretically on 72 cores i could do it in 40 hours
[08:50]   iris-discord2   <latenivenatrix9953> how does that compare to wrf
[08:53]   iris-discord2   <latenivenatrix9953> on 72 cores we get 72 hr/90 min
[08:54]   iris-discord2   <latenivenatrix9953> which is better than my laptop plasim speed but
[08:54]   iris-discord2   <latenivenatrix9953> that comes out to 7.5 days
[08:54]   iris-discord2   <latenivenatrix9953> and we can't reduce it cause it's already on 72 cores
[08:54]   iris-discord2   <latenivenatrix9953> so it's quite a bit faster
[08:55]   iris-discord2   <latenivenatrix9953> plasim is ~4.5x faster than wrf
[08:55]   iris-discord2   <latenivenatrix9953> with actually a comparable grid size
[08:56]   iris-discord2   <latenivenatrix9953> unfortunately my estimate was wildly off
[08:56]   iris-discord2   <latenivenatrix9953> i figured plasim could do this in about a day
[08:56]   iris-discord2   <latenivenatrix9953> on my laptop
[08:57]   iris-discord2   <latenivenatrix9953> hm

data & cost:
[09:24]   iris-discord2   <latenivenatrix9953> it seems to be producing 104MB of output every 6 hours
[09:24]   iris-discord2   <latenivenatrix9953> so a total of 202 GB for a year sim
[09:25]   iris-discord2   <latenivenatrix9953> that number will be important for determining how much space i need to pay for on aws
[09:28]   iris-discord2   <latenivenatrix9953> hm
[09:28]   iris-discord2   <latenivenatrix9953> if a 72 core thing is ~$1/hr
[09:28]   iris-discord2   <latenivenatrix9953> then 40 hours is $40, idk how much space is but it's probably not that much if i do all the simulations really quick like in a month or two
[09:29]   iris-discord2   <latenivenatrix9953> downloading the data on the other hand might be quite abit
[09:29]   iris-discord2   <latenivenatrix9953> but if i can process it fairly quickly (no idea if that's possible)
[09:29]   iris-discord2   <latenivenatrix9953> and into a nice low space state that shouldn't be that big of a deal either
[09:29]   iris-discord2   <latenivenatrix9953> let's just say it's $50 and double it for safety
[09:30]   iris-discord2   <latenivenatrix9953> and that means i can do a test run to make sure it works for about $100
[09:32]   iris-discord2   <latenivenatrix9953> then $100 for each successive run: a test with the parameters we want like seasonal cycles and testing to make sure the other things change, testing on earth as a control to see if cyclogenesis is reasonable, then any number of tests on pavala with different axial tilts and stuff so with one pavala run that's $400
[09:32]   iris-discord2   <latenivenatrix9953> if we want to test like 3 axial tilts like $600

it's better than wrf time and cost wise but it's still not ideal i'm not really sure what to do

we might have to go back to the genesis parameter on T21 or T42 (too low resolution for actually simulating cyclones)
« Last Edit: April 28, 2019, 02:57:51 AM by atomic7732 »