guidance on distributed bootstrap

[oricon]

I'm hoping to get some guidance on the best way to run parallel bootstraping on an HPC system. I have a complicated model - 233 coefficients with 32397 observations from 341 participants. To get an idea of the resources needed to run the bootstrapping I opened an interactive terminal with 48 cores and 360 GB of RAM. For 5,000 iterations the bootstrap gives me an estimate of 41 days. I wasn't sure if this was actually taking advantage of the multiple cores or not, so I tried to run it as distributed following the documentation:

using DataFrames, MixedModels, StatsModels, RData, CSV, Random, Distributed, ProgressMeter
addprocs(40)

# copy everything to workers
@showprogress for w in workers()
    remotecall_fetch(() -> coefnames(rt_fit3), w)
end

n_replicates = 10000
n_rep_per_worker = n_replicates ÷ nworkers()

pb_map = @showprogress pmap(MersenneTwister.(1:nworkers())) do rng
    parametricbootstrap(rng, n_rep_per_worker, rt_fit3; optsum_overrides)
end;

# remove the workers
for i in workers()
    rmprocs(i)
end

However, I get the following error after the @showprogress remotecall_fetch step:

julia> @showprogress for w in workers()
           remotecall_fetch(() -> coefnames(rt_fit3), w)
       end
ERROR: On worker 2:
UndefVarError: `coefnames` not defined
Stacktrace:
 [1] #19
   @ ./REPL[31]:2
 [2] #invokelatest#2
   @ ./essentials.jl:819 [inlined]
 [3] invokelatest
   @ ./essentials.jl:816
 [4] #110
   @ /sw/eb/sw/Julia/1.9.3-linux-x86_64/share/julia/stdlib/v1.9/Distributed/src/process_messages.jl:285
 [5] run_work_thunk
   @ /sw/eb/sw/Julia/1.9.3-linux-x86_64/share/julia/stdlib/v1.9/Distributed/src/process_messages.jl:70
 [6] macro expansion
   @ /sw/eb/sw/Julia/1.9.3-linux-x86_64/share/julia/stdlib/v1.9/Distributed/src/process_messages.jl:285 [inlined]
 [7] #109
   @ ./task.jl:514
Stacktrace:
 [1] remotecall_fetch(::Function, ::Distributed.Worker; kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
   @ Distributed /sw/eb/sw/Julia/1.9.3-linux-x86_64/share/julia/stdlib/v1.9/Distributed/src/remotecall.jl:465
 [2] remotecall_fetch(::Function, ::Distributed.Worker)
   @ Distributed /sw/eb/sw/Julia/1.9.3-linux-x86_64/share/julia/stdlib/v1.9/Distributed/src/remotecall.jl:454
 [3] #remotecall_fetch#162
   @ /sw/eb/sw/Julia/1.9.3-linux-x86_64/share/julia/stdlib/v1.9/Distributed/src/remotecall.jl:492 [inlined]
 [4] remotecall_fetch(::Function, ::Int64)
   @ Distributed /sw/eb/sw/Julia/1.9.3-linux-x86_64/share/julia/stdlib/v1.9/Distributed/src/remotecall.jl:492
 [5] macro expansion
   @ REPL[31]:2 [inlined]
 [6] top-level scope
   @ ~/.julia/packages/ProgressMeter/dMfiC/src/ProgressMeter.jl:1010

So I have 2 main questions:

1. Do I need to run bootstrapping as distributed on an HPC?
2. If so, is there an error in the documentation? If not, can I simply call more nodes to reduce the run time from 41 days?

[oricon]

I was able to figure this out. I needed to use @Everywhere to add the ProgressMeter and MixedModels modules to each worker. Add this code seems to have resolved my issue.

@everywhere begin
    using ProgressMeter, MixedModels
end

[palday]

Oh this should be fixed in the docs -- because the docs example only used one process, it didn't run into this issue in the same way. Sorry!