Eliminate runtime dispatch for repeated @async calls

General Usage Performance
1

Hi all!

I am writing some code where I need to call a function repeatedly, which is executed in a distributed fashion. Below you find some pseudocode to illustrate the problem. In outer_function I followed the suggestion in the manual (Shared Arrays Manual).

My Question: The profiling result suggests, that @async always needs runtime dispatch. Is there a way to get rid of the runtime dispatch?

Comments:

  • Running the entire code in parallel might be difficult because there are some changes which need to be done in the main outer loop before outer_function can be called.
  • I tried to get rid of the distributed part and just use multithreading, but the distributed solution is about twice as fast.
using Distributed
addprocs(15)
@everywhere using DistributedArrays, SharedArrays

function inner_function(result, parameters, data)
    local_ind = DistributedArrays.localpartindex(data)
    data_local = DistributedArrays.localpart(data)
    inner_result = 0  
    # Some more code, which includes a threaded for loop
    # and which will update inner_result 
    result[cind] = inner_result
    nothing
end

function outer_function(parameters::Vector{F}, data) where F
   # some more code here where something is done with parameters
   results = SharedArray{F, 1}((length(procs(data))), init=zero(F), pids=vec(procs(data)))
   @sync begin
        for j in procs(data)
             @async remotecall_wait(inner_function, j, results, parameters, data)
        end
   end
   # in the actual code this is a bit more elaborate
   return_value = sum(results) 
   return return_value
end

# prepare the data
data_raw = rand(10, 1000, 15)
data = distribute(data_raw, procs=workers(), dist=[1, 1, nworkers()]
parameters = rand(1000)
n_steps = 1000

# run the main loop
for step in 1:n_steps
    # do something to the parameters
    # ALL WORKERS NEED TO WAIT FOR THIS UPDATE TO BE DONE
    tmp_val = outer_function(parameters, data)
    # write tmp_val
end

Thank you so much for your help and suggestions!