Hi Julia! I am writing a small molecular dynamics engine in Julia to teach myself the language and I noticed that 67% of the time spent running the following Ewald summation function (to calculate coulomb energy of an infinite crystal lattice in reciprocal space) is spent in garbage collection. Bes…

style things: your type annotations aren’t helping size(x,1) is nicer than size(x)[1] for performance the biggest thing (I think) is that x[i, :] creates a copy, and if you want views (which you almost certainly do), you can put @views before the for loops.

[image] dkonstan: norm(xyz[i, :] - xyz[j, :] + L * ns[l, :]) Note that this makes quite a few arrays, not just the slices, but L * ns[l, :] is another, and -, and +. You can save a few by broadcasting these, .+ .* etc, or save all of them by not making slices. Maybe just sqrt(sum(abs2(xyz[i, …

[image] Oscar_Smith: your type annotations aren’t helping When you remove the type annotations be sure to replace strucFactor::ComplexF64 = 0.0 with strucFactor = zero(ComplexF64) to avoid type instability. exp(im * ...) can be replaced with cis(...) (abs(strucFactor)^2) cab be re…

It is easier to store coordinates as vectors of static vectors (from StaticArrays.jl). That will make the code cleaner and solve allocations that occur in operations like norm(x[i] - x[j]) where x[i or j] vectors of coordinates. ps. Take a look at http://github.com/m3g/CellListMap.jl for short rang…

[image] dkonstan: xyz[i, :] Am I guessing correctly that this is an nAtoms × 3 storing the xyz coordinates of atoms? Not only is this slice allocating a copy (slow), and is in the wrong order for locality (slow), but it’s also a lot faster to unroll such tiny loops over 3 components. All …

[image] stevengj: and is in the wrong order for locality (slow), I previously thought this was always the right answer, but Chris Elrod showed me how life is more complicated at times n = 256 a = randn(3, n) b = randn(n, 3) function optimal(a::Array{T}) where T s = zero(T) for i in …

In MD simulations one generally needs to compute only the distances of particles that are close to each other, and these are essentially random pairs of triplets of coordinates of the vectors. Thus the only locality that is guaranteed is that of the three coordinates of each particle, and the 3xN (o…

Thank you everyone! I will try your suggestions.

Ha, I am actually thinking of writing Ewald summation in julia myself. Perhaps a bit off-topic - is there any existing Ewald summation implementation in Julia already? As for the coding style, the previously answers are already very good. The key is to avoid slicing inside the loop, since julia def…

[image] zhubonan: The speed should be equivalent to: norm(xyz[i] - xyz[j] + L * ns[l]) where xyz and ns are Vector of static vectors It’s not equivalent, because (a) the loops/arrays you wrote are still in the wrong order for locality, (b) they may not be unrolled the way SVector operations…

Speeding up Ewald summation

New to Julia

jgreener64 November 3, 2025, 5:41pm 16

The latest release of Molly.jl has support for PME, see the post on this thread: Electrostatics in Julia - #33 by jgreener64.

Topic		Replies	Views
Electrostatics in Julia Modelling & Simulations	32	5137	November 3, 2025
optimising for loop performance Performance	14	3103	July 8, 2018
Performance improvements for simple molecular surface caclulation Performance	15	749	September 14, 2020
Can I make my code faster with parallelism, or just plain better coding? Performance	33	3008	February 19, 2019
Why does this Python code performs three times faster than Julia? Performance performance	22	1494	October 13, 2023

Speeding up Ewald summation

Related topics