What to do when you are stuck with Julia?

By: Blog by Bogumił Kamiński

Re-posted from: https://bkamins.github.io/julialang/2024/01/12/rgg.html

Introduction

Today I decided to write about code refactoring in Julia.
This is a topic that is, in my experience, a quite big advantage of this language.

A common situation you are faced with when writing your code is as follows.
You need some functionality in your program and it is available in some library.
However, what the library provides does not meet your expectations.
Since in Julia most packages are written in Julia under MIT license, it is easy to solve this issue.
You just take the source code and modify it.

Today I want to show you a practical example of such a situation I have had this week
when working with the Graphs.jl package.

The post was written using Julia 1.10.0, BenchmarkTools.jl 1.4.0, and Graphs.jl 1.9.0.

The problem

In my work I needed to generate random geometric graphs.
This is a simple random graph model that works as follows
(here I describe a general idea, for details please check the
Wikipedia entry on random geometric graphs).
To generate a graph on N vertices you first drop N random points
in some metric space. Next you connect two points with an edge
if their distance is less than some pre-specified distance.

The Graphs.jl library provides the euclidean_graph function
that generates such graphs. Here is a summary of its docstring:

euclidean_graph(N, d; rng=nothing, seed=nothing, L=1., p=2., cutoff=-1., bc=:open)

Generate N uniformly distributed points in the box [0,L]^{d}
and return a Euclidean graph, a map containing the distance on each
edge and a matrix with the points' positions.

An edge between vertices x[i] and x[j] is inserted if norm(x[i]-x[j], p) < cutoff.
In case of negative cutoff instead every edge is inserted.
Set bc=:periodic to impose periodic boundary conditions in the box [0,L]^d.

So what is the problem with this function? Unfortunately it is slow.
Let us, for example check how long it takes to compute an average degree
of a node in such a graph with n nodes and cutoff=sqrt(10/n), when setting
bc=:periodic (periodic boundary, i.e. distance is measured on a torus) for two-dimensional space.

julia> using Graphs

julia> for n in 1_000:1_000:10_000
           println(@time ne(euclidean_graph(n, 2; cutoff=sqrt(10/n), bc=:periodic, seed=1)[1])/n)
       end
  0.091657 seconds (2.50 M allocations: 193.170 MiB, 14.06% gc time)
15.604
  0.300285 seconds (10.00 M allocations: 765.801 MiB, 12.59% gc time)
15.661
  0.686230 seconds (22.50 M allocations: 1.686 GiB, 12.47% gc time)
15.744
  1.175881 seconds (40.00 M allocations: 2.990 GiB, 10.97% gc time)
15.7065
  1.800561 seconds (62.50 M allocations: 4.666 GiB, 10.76% gc time)
15.6568
  2.697535 seconds (90.00 M allocations: 6.716 GiB, 14.76% gc time)
15.641333333333334
  3.690599 seconds (122.50 M allocations: 9.138 GiB, 13.28% gc time)
15.743571428571428
  4.745701 seconds (160.00 M allocations: 11.932 GiB, 13.53% gc time)
15.714
  5.962431 seconds (202.51 M allocations: 15.099 GiB, 12.70% gc time)
15.722222222222221
  7.195257 seconds (250.01 M allocations: 18.638 GiB, 11.42% gc time)
15.7086

We can see that the euclidean_graph function scales badly with n.
Note that by choosing cutoff=sqrt(10/n) we have a roughly constant
average degree, so the number of edges generates scales linearly, but
the generation time seems to grow much faster.

The investigation

To find out the source of the problem we can investigate the source
of euclidean_graph, which consists of two methods:

function euclidean_graph(
    N::Int,
    d::Int;
    L=1.0,
    rng::Union{Nothing,AbstractRNG}=nothing,
    seed::Union{Nothing,Integer}=nothing,
    kws...,
)
    rng = rng_from_rng_or_seed(rng, seed)
    points = rmul!(rand(rng, d, N), L)
    return (euclidean_graph(points; L=L, kws...)..., points)
end

function euclidean_graph(points::Matrix; L=1.0, p=2.0, cutoff=-1.0, bc=:open)
    d, N = size(points)
    weights = Dict{SimpleEdge{Int},Float64}()
    cutoff < 0.0 && (cutoff = typemax(Float64))
    if bc == :periodic
        maximum(points) > L && throw(
            DomainError(maximum(points), "Some points are outside the box of size $L")
        )
    end
    for i in 1:N
        for j in (i + 1):N
            if bc == :open
                Δ = points[:, i] - points[:, j]
            elseif bc == :periodic
                Δ = abs.(points[:, i] - points[:, j])
                Δ = min.(L .- Δ, Δ)
            else
                throw(ArgumentError("$bc is not a valid boundary condition"))
            end
            dist = norm(Δ, p)
            if dist < cutoff
                e = SimpleEdge(i, j)
                weights[e] = dist
            end
        end
    end
    g = Graphs.SimpleGraphs._SimpleGraphFromIterator(keys(weights), Int)
    if nv(g) < N
        add_vertices!(g, N - nv(g))
    end
    return g, weights
end

The beauty of Julia is that this source is written in Julia and is pretty short.
It immediately allows us to pinpoint the source of our problems. The core of we work
is done in a double-loop iterating over i and j indices. So the complexity of this
algorithm is quadratic in number of vertices.

Fixing the problem

The second beauty of Julia is that we can easily fix this. The idea can be found in
the Wikipedia entry on random geometric graphs in the algorithms section
here.

A simple way to improve the performance of the algorithm is to notice that if
you know L and cutoff you can partition the space into equal sized cells
having side length floor(Int, L / cutoff). Now you see that if you have a vertex
in some cell then it can be connected only to nodes in the same cell or cells directly
adjacent to it (the cells that are more far away contain the points that must be farther
than cutoff from our point). This means that we will have a much lower number of points
to consider. Below I show a code that is a modification of the original source that
adds this feature. The key added function is to_buckets which computes the bucket
identifier for each vertex and creates a dictionary that is a mapping from bucked
identifier to a vector of node numbers that fall into it:

using LinearAlgebra
using Random

function euclidean_graph2(
    N::Int,
    d::Int;
    L=1.0,
    rng::Union{Nothing,AbstractRNG}=nothing,
    seed::Union{Nothing,Integer}=nothing,
    kws...,
)
    rng = Graphs.rng_from_rng_or_seed(rng, seed)
    points = rmul!(rand(rng, d, N), L)
    return (euclidean_graph2(points; L=L, kws...)..., points)
end

function to_buckets(points::Matrix, L, cutoff)
    d, N = size(points)
    dimlen = max(floor(Int, L / max(cutoff, eps())), 1)
    buckets = Dict{Vector{Int}, Vector{Int}}()
    for (i, point) in enumerate(eachcol(points))
        bucket = floor.(Int, point .* dimlen ./ L)
        push!(get!(() -> Int[], buckets, bucket), i)
    end
    return buckets, dimlen
end

function euclidean_graph2(points::Matrix; L=1.0, p=2.0, cutoff=-1.0, bc=:open)
    d, N = size(points)
    weights = Dict{Graphs.SimpleEdge{Int},Float64}()
    cutoff < 0.0 && (cutoff = typemax(Float64))
    if bc == :periodic
        maximum(points) > L && throw(
            DomainError(maximum(points), "Some points are outside the box of size $L")
        )
    end
    buckets, dimlen = to_buckets(points, L, cutoff)
    deltas = collect(Iterators.product((-1:1 for _ in 1:size(points, 1))...))
    void = Int[]
    for (k1, v1) in pairs(buckets)
        for i in v1
            for d in deltas
                k2 = bc == :periodic ? mod.(k1 .+ d, dimlen) : k2 = k1 .+ d
                v2 = get(buckets, k2, void)
                for j in v2
                    i < j || continue
                    if bc == :open
                        Δ = points[:, i] - points[:, j]
                    elseif bc == :periodic
                        Δ = abs.(points[:, i] - points[:, j])
                        Δ = min.(L .- Δ, Δ)
                    else
                        throw(ArgumentError("$bc is not a valid boundary condition"))
                    end
                    dist = norm(Δ, p)
                    if dist < cutoff
                        e = Graphs.SimpleEdge(i, j)
                        weights[e] = dist
                    end
                end
            end
        end
    end

    g = Graphs.SimpleGraphs._SimpleGraphFromIterator(keys(weights), Int)
    if nv(g) < N
        add_vertices!(g, N - nv(g))
    end
    return g, weights
end

Note that it took less than 30 lines of code to add the requested feature to the code.

Performance of the improved code

Let us test our new code:

julia> for n in 1_000:1_000:10_000
           println(@time ne(euclidean_graph2(n, 2; cutoff=sqrt(10/n), bc=:periodic, seed=1)[1])/n)
       end
  0.017221 seconds (274.10 k allocations: 21.751 MiB, 22.15% gc time)
15.604
  0.022855 seconds (558.52 k allocations: 42.289 MiB, 10.43% gc time)
15.661
  0.032693 seconds (852.46 k allocations: 69.684 MiB, 8.52% gc time)
15.744
  0.043141 seconds (1.10 M allocations: 87.196 MiB, 14.73% gc time)
15.7065
  0.071273 seconds (1.41 M allocations: 109.725 MiB, 7.67% gc time)
15.6568
  0.068194 seconds (1.70 M allocations: 130.828 MiB, 12.54% gc time)
15.641333333333334
  0.071277 seconds (1.98 M allocations: 150.712 MiB, 11.85% gc time)
15.743571428571428
  0.081463 seconds (2.24 M allocations: 169.153 MiB, 10.67% gc time)
15.714
  0.099957 seconds (2.48 M allocations: 186.492 MiB, 8.08% gc time)
15.722222222222221
  0.148573 seconds (2.84 M allocations: 213.214 MiB, 18.37% gc time)
15.7086

We seem to get what we wanted. Our computation time looks to scale quite well.
Also the obtained average degree numbers are identical to the original ones.

Let us compare the performance on an even larger graph:

julia> n = 100_000;

julia> @time ne(euclidean_graph(n, 2; cutoff=sqrt(10/n), bc=:periodic, seed=1)[1])/n
908.976252 seconds (25.00 G allocations: 1.819 TiB, 11.64% gc time, 0.00% compilation time)
15.70797

julia> julia> @time ne(euclidean_graph2(n, 2; cutoff=sqrt(10/n), bc=:periodic, seed=1)[1])/n
  1.640495 seconds (27.53 M allocations: 2.121 GiB, 19.83% gc time)
15.70797

Indeed we see that the timing of the original implementation becomes prohibitive for
larger graphs.

Making sure things are correct

Before we finish there is one important task we need to make. We should check if
indeed the euclidean_graph2 function produces the same results as euclidean_graph.
This is easy to test with the following randomized test:

julia> using Test

julia> Random.seed!(1234);

julia> @time for i in 1:1000
           N = rand(10:500)
           d = rand(1:5)
           L = rand()
           p = 3*rand()
           cutoff = rand() * L / 4
           bc = rand([:open, :periodic])
           seed = rand(UInt32)
           @test euclidean_graph(N, d; L, p, cutoff, bc, seed) ==
                 euclidean_graph2(N, d; L, p, cutoff, bc, seed)
       end
 16.955773 seconds (275.09 M allocations: 20.342 GiB, 12.27% gc time)

We have tested 1000 random setups of the experiments. In each of them both functions
returned the same results.

Conclusions

In my post I have shown you an example that one can easily tweak some package code to your needs.
In this case this was performance, but it can be equally well functionality.

I did not comment too much on the code itself, as it was a bit longer than usual, but let me
discuss as a closing remark one performance aspect of the code. In my to_buckets function I used
the get! function to populate the dictionary with a mutable default value (Int[] in that case).
You might wonder why I preferred to use an anonymous function instead of passing a default
as a third argument. The reason is number of allocations. Check this code:

julia> using BenchmarkTools

julia> function f1()
           d = Dict(1 => Int[])
           for i in 1:10^6
               get!(d, 1, Int[])
           end
           return d
       end
f1 (generic function with 1 method)

julia> function f2()
           d = Dict(1 => Int[])
           for i in 1:10^6
               get!(() -> Int[], d, 1)
           end
           return d
       end
f2 (generic function with 1 method)

julia> @benchmark f1()
BenchmarkTools.Trial: 195 samples with 1 evaluation.
 Range (min … max):  19.961 ms … 45.328 ms  ┊ GC (min … max): 10.26% … 13.19%
 Time  (median):     23.962 ms              ┊ GC (median):    10.45%
 Time  (mean ± σ):   25.660 ms ±  5.050 ms  ┊ GC (mean ± σ):  12.27% ±  3.76%

    ▃▂▂█▃▃ ▃▂▃
  ▆▄██████▇███▇▆▄▄▇▄▄▁▃▆▄▃▅▄▅▄▄▄▃▄▃▁▃▃▁▁▁▃▃▃▃▃▃▁▃▁▁▃▁▁▁▁▁▃▁▁▃ ▃
  20 ms           Histogram: frequency by time        44.3 ms <

 Memory estimate: 61.04 MiB, allocs estimate: 1000005.

julia> @benchmark f2()
BenchmarkTools.Trial: 902 samples with 1 evaluation.
 Range (min … max):  4.564 ms … 11.178 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     5.149 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   5.526 ms ±  1.396 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

  █▅▄▄▃▃▄▅▆▄▁▁                                           ▃
  ██████████████▆▆▆▆▆▅▇▆▄▆▆▅▅▁▄▅▁▅▁▇▅▆▄▄▇▁▅▅▅▄▁▄▄▆▅▁▅▁▁▅▆██▅ █
  4.56 ms      Histogram: log(frequency) by time       10 ms <

 Memory estimate: 592 bytes, allocs estimate: 5.

As you can see f2 is much faster than f1 and does much less allocations. The issue
is that f1 allocates a fresh Int[] object in every iteration of the loop, while
f2 would allocate it only if get! does not hit a key that already exists in d
(and in our experiment I always queried for 1 which was present in d).