Julia: How to execute functions in parallel? - parallel-processing

Question

I want to run functions in parallel. These functions are executed many times in a loop.
coordSys = SharedArray{Bool}([true,false,true,true]);
dir = SharedArray{Int8}([1,2,3,2]);
load = SharedArray{Float64}([8,-7.5,7,-8.5]);
L = SharedArray{Float64}([400,450,600,500]);
r = SharedArray{Float64}([0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0]);
Obviously these vectors will be huge, but for simplicity I just put this limited size.
Operation without parallel computing:
function unifLoad(coordSys,dir,load,L,ri)
if coordSys == true
if dir == 1
Q = [load;0;0];
elseif dir == 2
Q = [0;load;0];
elseif dir == 3
Q = [0;0;load];
end
q = ri*Q; #matrix multiplication
P = q[1]*L/2;
V = q[2]*L/2;
M = -q[3]*L*L/12;
f = [P;V;M];
else
f = [1.0;1.0;1.0];
end
return f
end
running the loop:
var = zeros(12)
for i = 1:length(L)
var[3*(i-1)+1:3*i] = unifLoad(coordSys[i],dir[i],load[i],L[i],r[3*(i-1)+1:3*i,:]);
end
The returned value is:
var
12-element Array{Float64,1}:
0.0
0.0
-1.06667e5
1.0
1.0
1.0
2100.0
0.0
-0.0
0.0
2125.0
-0.0
Operation with parallel computing
I've been trying to implement the same function in parallel, but without getting the same results.
# addprocs(3)
#everywhere function unifLoad_Parallel(coordSys,dir,load,L,ri)
if coordSys == true
if dir == 1
Q = [load;0;0];
elseif dir == 2
Q = [0;load;0];
elseif dir == 3
Q = [0;0;load];
end
q = ri*Q; # Matrix multiplication (ri -> Array 3x3)
P = q[1]*L/2;
V = q[2]*L/2;
M = -q[3]*L*L/12;
f = [P;V;M];
else
f = [1.0;1.0;1.0];
end
return f
end
running the parallel loop:
var_parallel = SharedArray{Float64}(12);
#parallel for i = 1:length(L)
var_parallel[3*(i-1)+1:3*i] = unifLoad_Parallel(coordSys[i],dir[i],load[i],L[i],r[3*(i-1)+1:3*i,:]);
end
The returned value is:
var_parallel
12-element SharedArray{Float64,1}:
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

Answer 1

On my Julia 0.6.3 the parallel code returns the same result so I am unable to reproduce the problem (also I do not encounter the issue #SalchiPapa reports).
However, I would like to note that this code actually should work faster with threads (I assume that the real problem is much larger). Here is the code you could use (I used an equivalent implementation to your which is a bit shorter - but the only significantly relevant change is that I wrap it in a function which provides dramatic performance gains). The crucial issue that all arrays except var are shared but only read. And var is written but only once at each entry and not read from. This is the case where it is safe to use threading which has a lower overhead.
Here is an example code (you have to define JULIA_NUM_TREADS environment variable before starting Julia and set it to number of threads you want - most probably 4 is what you want):
using Base.Threads
function experiment()
coordSys = [true,false,true,true];
dir = [1,2,3,2];
load = [8,-7.5,7,-8.5];
L = [400,450,600,500];
r = [0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0];
unifLoad(coordSys,dir,load,L,r, i) =
coordSys ? load * L * r[3*(i-1)+1:3*i, dir] .* [0.5, 0.5, -L/12] : [1.0, 1.0, 1.0]
var = zeros(12)
#threads for i = 1:length(L)
var[3*(i-1)+1:3*i] = unifLoad(coordSys[i],dir[i],load[i],L[i],r,i);
end
var
end
Also here is a bit simplified code for parallel processing using similar ideas:
coordSys = SharedArray{Bool}([true,false,true,true]);
dir = SharedArray{Int8}([1,2,3,2]);
load = SharedArray{Float64}([8,-7.5,7,-8.5]);
L = SharedArray{Float64}([400,450,600,500]);
r = SharedArray{Float64}([0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0
0.0 0.0 1.0; 0.0 -1.0 0.0; 1.0 0.0 0.0]);
#everywhere unifLoad(coordSys,dir,load,L,r,i) =
coordSys ? load * L * r[3*(i-1)+1:3*i, dir] .* [0.5, 0.5, -L/12] : [1.0, 1.0, 1.0]
vcat(pmap(i -> unifLoad(coordSys[i],dir[i],load[i],L[i],r,i), 1:length(L))...)
Here pmap is mostly used to simplify the code so that you do not need #sync.