Variable declaration and initialisation

New to Julia
1

I am trying to optimise a piece of my code. I have currently used mutable struct to declare variable types. Following which I assign values to them,

mutable struct Par
    N::Int64 ; 
    n::Int64 ; 
end
mutable struct Mat
    A::Array{Int64,2} ; 
    θₜ::Array{Float64,2} ; 
    Atemp::Array{Int64,2} ; 
    Asep::Array{Float64,2} ;
    allag::Array{Float64,2} ; 
end
p = Par(1000,5000) ;
m = Mat(zeros(Int64,p.N,p.N),zeros(Float64,p.n,p.N),zeros(Int64,p.N,p.N),zeros(Float64,p.N,p.N),zeros(Float64,p.N,p.N)) ;

I would want to increase values of N and n in subsequent simulations. But I notice already that initialising the matrices themselves take time of the order of 10^-3,

using BenchmarkTools
@btime m = Mat(zeros(Int64,p.N,p.N),zeros(Float64,p.n,p.N),zeros(Int64,p.N,p.N),zeros(Float64,p.N,p.N),zeros(Float64,p.N,p.N)) ;
  13.360 ms (21 allocations: 68.67 MiB)

Is there a faster/optimal way of doing this?

2

do you need to initialize different matrix of this size millions of times? (potentially you can use Undef) Looking at the allocation size, I think whatever you do you will be memory bounded.

3

I am carrying out MonteCarlo simulations, and I would like to increase N from 1000 to 100,000 or so. I was trying to see if I could save some time here to better the execution time.

4

I would try to design the algorithm to use a buffer pre-allocated, and reused over and over again.

1 Like
5

Can you please suggest how to do so?

6

How about running the MC in parallel. Let us say in threads. Then each thread would have a buffer of the type you specified above, and when it starts it overwrites the values in the buffer with the correct initial values and takes off. This way you need to allocate only T thread buffers. Obviously, because each thread only accesses its own buffer, there is no memory race…And this way you don’t run out of memory when allocating lots of matrices that get used only once :slight_smile:

1 Like
7

Thank you. I am relatively new to Julia and not an ace programmer. Can I ask if there is any documentation on the use of threads, some example to look up to?

8 
using Base.Threads
struct ThreadBuffer
	t::Int64
	parameters::Vector{Float64}
end

@show parameters = rand(10)
nth = 3
chunk = Int64(round(length(parameters) / nth))
buffers = ThreadBuffer[]
for t in 1:nth-1
	push!(buffers, ThreadBuffer(t, parameters[chunk*(t-1)+1:chunk*(t)+1-1]))
end
push!(buffers, ThreadBuffer(nth, parameters[chunk*(nth-1)+1:length(parameters)]))
@show buffers

threadwork(t) = begin
for p in 1:length(buffers[t].parameters)
	println("Thread $(t), Working with parameter = $(buffers[t].parameters[p])")
	sleep((nth-t+1)*rand())
end
end

tasks = []
Threads.@threads for t in 1:nth
	push!(tasks, Threads.@spawn threadwork(t))
end
for t in 1:length(tasks)
	Threads.wait(tasks[t])
end
println("Done")
3 Likes
9

I’d strongly suggest getting the serial version of your code working
before worrying about multi-threading. The only gotcha when moving
from serial to parallel is making sure that any functions using random
numbers take an rng as an argument rather than just using the global
rng.

What we mean when we say use a pre-allocated buffer is just to do
something like the following for memory of size n*n with m
replications:

# Pre-allocate your matrix/vector/chunk of memory
mem = Matrix{Float64}(undef, n, n)
results = Vector{Float64}(undef, n)
for i in eachindex(results)
    for j in eachindex(mem)
        mem[j] = # do calculation
    end#for
    results[i] = f(mem)
end

Basically just re-use arrays you allocate as much as possible. If n
varies between runs of your simulation, you can still re-use the
arrays: just change them to the appropriate size with resize! (under
the hood, julia may have to reallocate the array and copy its values
to a different location in memory when resize!ing, but user-code
doesn’t have to worry about that).

3 Likes
10

Thank you. I already have the Serial version of my code working. Since, it takes really long, I am trying ways to optimise it and better the execution time. Any suggestions for doing the same will be of great help.

11

I was trying to understand what each syntax in this code does. When I tried running,

import Base.Threads.@spawn

Threads.@threads for t in 1:nth
	push!(tasks, Threads.@spawn threadwork(t))
end

I get the following errors,

WARNING: could not import Threads.@spawn into Main
UndefVarError: @spawn not defined

Is there something else that needs to be done?

12

Please remember to run this with Julia 1.3.

13

That worked. Thank you.

14

If the post by @PetrKryslUCSD answered your question it is a good idea to mark it as “Solution”.

1 Like