RAM usage of building a system image

ufechner7 · October 14, 2026, 8:26pm

My script that is building a system image needs about 50 GB of RAM. This is very expensive. I am trying to use a swap file, but this is getting very slow.
See for example: Install KiteModels#main · OpenSourceAWE/AWEMeta.jl@2a4ca15 · GitHub

Did not finish in 4h, now trying 6h.

Any ideas how to reduce the RAM usage? I switched from ControlPlots.jl to MakieControlPlots.jl. Before it was kind of OK.

You can find the script that builds the system image here:

github.com/OpenSourceAWE/KiteModels.jl

test/create_sys_image.jl

main

# SPDX-FileCopyrightText: 2022, 2024, 2025 Uwe Fechner
# SPDX-License-Identifier: MIT

# activate the test environment if needed
using Pkg
if dirname(Pkg.project().path) != @__DIR__
    Pkg.activate(@__DIR__)
    Pkg.resolve()
    Pkg.instantiate()
end
@info "Loading packages ..."
using AtmosphericModels, Colors, ControlSystemsBase, DSP, Dierckx, DiscretePIDs,
    DocStringExtensions, JLD2, KitePodModels, KiteUtils, KiteViewers, LinearAlgebra, NLsolve,
    NonlinearSolve, OrdinaryDiffEqBDF, OrdinaryDiffEqCore, OrdinaryDiffEqNonlinearSolve,
    Parameters, REPL, Rotations, StaticArrays, StatsBase, Sundials, WinchModels
using BenchmarkTools, Documenter, PackageCompiler

@info "Creating sysimage ..."
push!(LOAD_PATH,joinpath(pwd(),"src"))

This file has been truncated. show original

Do you think using JULIA_HEAP_SIZE_HINT could help?

kristoffer.carlsson · October 15, 2026, 11:57am

You could check out aotcompile: Release the Julia heap before compiling (let's use all the threads) - Pull Request #63531 - JuliaLang/julia - GitHub

ufechner7 · October 17, 2026, 5:28am

Nice that things will improve with Julia 1.14, but why this large regression from Julia 1.12 to Julia 1.13: Building a system image using Julia 1.13 fails (due to excessive RAM usage) · Issue #63141 · JuliaLang/julia · GitHub

Can’t this regression be adressed?

ufechner7 · October 17, 2026, 2:06pm

A closer investigation revealed that it is not a Julia regression, but a PackageCompiler regression:

Since #1096 (released in PackageCompiler 2.4.1), create_sysimage defaults to cpu_target="sysimage" on Julia 1.13+, while older Julia versions keep "native". With the official Julia binaries this expands to

generic;sandybridge,-xsaveopt,clone_all;haswell,-rdrnd,base(1);x86-64-v4,-rdrnd,base(1)

so every function is compiled at least twice, plus partial clones for two more targets. For large sysimages this makes the “compiling incremental system image” step need more than twice the memory and several times as long.

Workaround: Use cpu_target="native". This fixed the issue for me.

New issue in PackageCompiler: Default `cpu_target="sysimage"` on Julia 1.13 more than doubles sysimage build memory and time · Issue #1143 · JuliaLang/PackageCompiler.jl · GitHub

Topic		Replies	Views
Sysimage creation process gets OOM killed (out-of-memory) Performance memory , package-compiler , sysimage	3	229	March 7, 2026
Creating system images much faster with Julia 1.10.0-beta1 General Usage package , package-compiler	8	1239	July 30, 2023
Creating a system image on Raspery Pi 5 General Usage package-compiler , modelingtoolkit	2	617	December 25, 2023
PackageCompiler crash after 6 hours of compilation, seemingly running out of memory General Usage package , package-compiler	3	540	July 7, 2024
Building a system image using Julia 1.13 fails on Windows General Usage question , package , package-compiler	1	188	October 20, 2026

RAM usage of building a system image

Related topics