I have these three ways of creating a zero vector. module m function zeroout(a) @inbounds Threads.@threads for k in eachindex(a) a[k] = 0.0 end end N = 1000000000 println("a = fill(0.0, N);") @time a = fill(0.0, N); println("undefined, a .= 0.0;") @time a = Vector{Float64}(undef,…

They have optimized implementations that branch on sizes, for example. They’ll use noon-temporal stores when the memory is too big to fit in cache, for example. It should in general be the fastest way to set memory (but calloc should still be better for 0).

[image] PetrKryslUCSD: Is there something faster? You could use calloc. See e.g. Faster zeros with calloc and ArrayAllocators.jl

I do need to write into it later. So it looks like the threaded version is the fastest overall.

[image] PetrKryslUCSD: I do need to write into it later. You can write into a calloc-allocated array.

Correct. But at that point the OS will do any delayed zeroing out, and things will slow down.

That’s a pretty surprising conclusion; I’d think your benchmarks aren’t as representative as you’d want them to be before coming to that conclusion. As they’re written there, you’re likely intermingling GC and compilation times.

@mbauman You are right. I had to be a bit more sophisticated. I obtained pkrysl@firenze:~/testheat$ ../julia-1.10.1/bin/julia -t 4 m.jl a = fill(0.0, N); 4.156 s (2 allocations: 7.45 GiB) a = zeros(N); ; 4.156 s (2 allocations: 7.45 GiB) undefined, a .= 0.0; 108.077 μs (2 allocations: 7.45 GiB…

The multithreaded function for zeroing out the vector pays off even for moderately long (short!) vectors. For instance ../julia-1.10.1/bin/julia -t 4 m.jl N = 100000 = 100000 a = fill(0.0, N); 76.781 μs (2 allocations: 781.30 KiB) a = zeros(N); 76.991 μs (2 allocations: 781.30 KiB) undefined, …

[image] PetrKryslUCSD: But at that point the OS will do any delayed zeroing out, and things will slow down. It is possible that the OS or libc implementation already has a pool of zeroed memory pages that it can just assign.

I’m afraid expecting that to happen will not be very robust, will it?

You can rely on the OS providing zeroed out memory. At least in some or all modern OSes, if not it’s a security risk, since you read data from other processes. But most allocations you do (except at the start of your program) not come from the OS. They are you own memory reused. It doesn’t seem ver…

Fastest way of getting a long zero vector?

General Usage

stevengj March 14, 2024, 6:01pm 2

You could use calloc. See e.g. Faster zeros with calloc and ArrayAllocators.jl

3 Likes

Topic		Replies	Views
Faster zeros with calloc Performance array	26	4090	March 23, 2022
[ANN] ArrayAllocators.jl v0.3 composes with OffsetArrays.jl v1.12.1+ for faster zeros with offset indexing Package Announcements announcement	4	931	June 30, 2022
Could fill! be twice as fast? Internals & Design	13	2382	May 15, 2017
Performance of resize! vs pre-allocating with zeros(...) Performance	9	1198	April 29, 2022
What about an `undefs` function in `Base`? General Usage feature-request	64	3512	January 24, 2023

Fastest way of getting a long zero vector?

Related topics