[julia-users] No performance gain from pmap

[Gabriel Mihalache]

Hello!
I would greatly appreciate any thought on these findings:

I run my code in 3 different cases, and the performance is pretty much the 
same:
- 1 node, 2 cores on it
- 1 node, 24 cores on it
- 3 nodes, 24 cores on each

I made sure that
1. the function is indeed being run on the right host/worker (using 
println-s)
2. the function has the right values for the various variables it needs

The issue seems to be that all the julia processes other than the master 
one are effectively idle. This is the top output under various scenarios, 
on various machines (master or not): https://imgur.com/a/5gtL7

This is how I add workers:

hostname = chomp(readall(`hostname`))
if hostname == ARGS[2]
        println("$hostname master")
        tasksPerNode = parseint(ARGS[1])

        nodes = ARGS[2:end]

        if length(nodes) > 1
                for ix = 1:tasksPerNode
                        addprocs(nodes; dir="/software/julia/0.3.2/bin/")
                end
        else
                addprocs(tasksPerNode)
        end

        @everywhere cd("/scratch/gmihalac/exitSwap/")
        @everywhere include("swap.jl")
        ExitSwap.compute()
else
        println("$hostname not master")
end


I am attaching the swap.jl file, to show how I call/use pmap.

I'll try to put together a small example, to see if i can replicate this in 
a simpler setup.

Any comments/ideas/suggestions are greatly appreciated!

Thank you!
Gabriel

swap.jl
Description: Binary data