I want to use threading to build a result. But it seems push! isn’t thread-safe: num_monte = Int(1e6) solution_data = Vector{Float64}() Threads.@threads for i in 1:num_monte push!(solution_data,i) end With @parallel there is a reduction, which I can use (vcat). Is there some …

num_monte = Int(1e6) solution_data = Vector{Vector{Float64}}() for i in 1:Threads.nthreads() push!(solution_data,Float64[]) end Threads.@threads for i in 1:num_monte push!(solution_data[Threads.threadid()],i) end vcat(solution_data...) julia> vcat(solution_data...) 1000000-element Array{Float64…

If you know the total size in advance, it would probably be faster to allocate the entire vector at the beginning, and then precalculate offsets into the vector for each thread, with each thread setting a separate section, so that no locking is needed, and no concatenation is needed at the end.

Is accessing an array like that thread safe? I do know the total size in advance.

The number of threads isn’t necessarily a divisor of the number of steps in the loop. How does @threads make the decision for how much work goes to each thread? If I do something fancy, won’t I need to match that exactly?

Yes, the code would be more complex to deal with having one of the threads handle more steps (the remainder after dividing it up). It would probably only be worth the effort if the array were pretty large. You wouldn’t be able to use the @threads macro I think.

[image] ChrisRackauckas: for i in 1:Threads.nthreads() push!(solution_data,Float64) end Note that there’ll likely be a small improvement if you do. solution_data = Vector{Vector{Float64}}(Threads.nthreads()) @threads for i in 1:Threads.nthreads() solution_data[i] = Float64[] end Due to …

Note: I don’t think that first part will actually improve performance, since you will be conflicting between each thread when you are just setting up a small number of entries in solution_data. (they will probably all hit the same cache line).

Sorry for reviving this old thread; I wanted to ask what exactly the intended behaviour of resizing/reallocating ops on arrays is, in multithreaded. As far as I understood the source (please confirm or correct), we have the following race behaviour: Bitstype elements: One thread causes realloc (…

I think that’s the wrong way to look at it. Instead, this is just how the computer actually works and the possibly extra safety features just don’t exist yet to abstract it away

I’m not sure I understand you here. Thread-safety often comes with a performance big cost, so these are not necessarily “not yet” safety features, but might instead be “it’s a feature, not a bug”. If I want to figure out how thread-unsafe array operations are, in practice, then I should read array.…

Thread-safe array building

General Usage

yuyichao April 19, 2017, 3:48am 2

Create nthreads() number of arrays and push to them based on threadid().

Safe loop with push! multi-threading

@threads in for loop and push! in if statement

Safe loop with push! multi-threading

Multi-threaded Array building

Topic		Replies	Views
Multi-threaded Array building Performance multithreading	22	516	September 20, 2025
Why is push! still so slow? Internals & Design	39	2068	March 29, 2024
Use Dynamic Arrays (Vectors) during Multi-threading Julia at Scale multithreading , array , memory-allocation	7	2120	September 10, 2018
Safe loop with push! multi-threading Performance performance , parallel	7	4195	June 23, 2020
Julia multithreading New to Julia question	8	616	January 15, 2023

Thread-safe array building

Related topics