Building a PC optimized for "time to first plot"

This is a bit unfair. Setting this up is not that trivial, compared to other languages, especially when working with new students. Sure, it is 3 or 4 steps that are trivial for me, but they require that you have learnt the julia idioms in order to understand the steps (Package.toml, sysimage, module, precompile, SnoopPrecompile, etc). You can blindly follow the steps without knowing the idioms, but if that is the approach a student takes, at the first innocuous error message or typo they will crash and burn.

It is not the process that is difficult, it is wrapping your head around why the process is necessary that is the largest stumbling block for people new to julia. We do need better tooling before we can be so dismissive (and I know that you and many other volunteers are making this happen and I immensely appreciate it – the progress over the last couple of years has been monumental).


That’s a bit unfair. With the new SnoopPrecompile mechanism, it’s fairly trivial.

Step 1: PkgTemplates:
Step 2: Change the module file to:

module MyLocalModule

import SnoopPrecompile
import # other packages

SnoopPrecompile.@precompile_all_calls begin
  # put code here


Step 3: ]add packages.

Step 4:

using PackageCompiler
PackageCompiler.create_sysimage(["MyLocalModule"]; sysimage_path="")

You’re done. It takes less than 5 minutes (of “you time”, the sysimage build can take longer). I do this on the daily. Training undergrads is sufficiently done just by sending a youtube video (Developing Julia Packages - YouTube). I mean, there are things to improve but let’s not act like this piece is hard.

It could be made into one click in VS Code (and that would be nice), but for now, someone who’s an experienced Julia developer (like the OP) can definitely do those few steps I mentioned.


I work with talented undergrads at the same institution as you. This statement is just not true. I, from the start, agree with you that this is trivial for developers with a couple of weeks of experience in Julia, but in my first post I was talking about students. Fair point if that is our point of disconnect, as that was my mindset when I complained about your “which apparantely no one does” comment.

And let me reiterate: the steps are trivial, but if the student is supposed to not crash and burn at the first typo, they need to understand the reasons for the steps. That requires learning quite a few new idioms (which is certainly worthwhile and they should do it, but it is not trivial).


Taking “time to first plot” literally, has anyone done snooping for Plots, or would that not work?

Or taking the question as being about hardware, an M1 mac gives these times (Julia nightly, native, Plots v1.35.3, no system image or any other special tricks). Might be interesting to see other systems, my impression is that apple silicon is pretty good for this:

% julia --startup-file=no -e '@time using Plots; @time display(plot(rand(20))); exit()'
  3.658917 seconds (6.30 M allocations: 435.703 MiB, 5.13% gc time, 19.33% compilation time: 54% of which was recompilation)
  3.703820 seconds (403.54 k allocations: 20.986 MiB, 99.42% compilation time: <1% of which was recompilation)
1 Like

Well I’ve found by homework 1 the undergrads are fine turning in problems as a tested package, so :person_shrugging:. But anyways, the OP is an experienced Julia developer who was with Julia Computing for quite some time, so that’s a bit of a shift in the topic. I think someone who has been developing Julia packages at Julia Computing should, with 2022 tooling, be able to comfortably build a system image.

Well I think in 2022, every package should be properly SnoopPrecompiled and everyone should be building a system image which includes calls that they use.

Indeed, in earlier times, system images didn’t do very much because packages were not setup to work well with precompilation (nothing was snooped, and pre-v1.8 changes that are detailed in the blog post). But if packages are setup properly, then you don’t even need to give a custom script and it will precompile what you need. With that in place, people should really start using system images more often.

For example, if you don’t develop static arrays, slap it into the system image. If you don’t develop the ODE solvers, slap OrdinaryDiffEq.jl into the system image. For most developers, a large part of their dependency stack can be relatively static.


I am not sure whether I am interpreting your question correctly, but I think the answer is yes. Here is an example with Makie:

With a sysimage that took just a couple of clicks to prepare in vscode:

julia> @time @eval using CairoMakie
  0.008680 seconds (14.54 k allocations: 927.412 KiB)

julia> @time plot([1,2,3])
  2.082096 seconds (1.94 M allocations: 130.274 MiB, 2.15% gc time, 37.41% compilation time)

Without a custom sysimage:

julia> @time @eval using CairoMakie
 13.082514 seconds (17.49 M allocations: 1.247 GiB, 4.57% gc time, 3.01% compilation time: <1% of which was recompilation)

# don't even bother actually making the first plot, it takes more than 30 seconds

Plots has a lot of dynamism so with current precompilation a lot still is dropped (though IIUC there is work going in on that), but it is already snooped:

Because it’s snooped, if you do:

using PackageCompiler
PackageCompiler.create_sysimage(["Plots"]; sysimage_path="")

then that system image is sufficient for first time GR plots to be quite fast (<0.1 seconds IIRC). Creating a sysimage for fast plotting with Plots.jl · PackageCompiler showcases a custom precompile file and all of that, but after the snooping was added it’s kind of not necessary.

1 Like

Ah great, I didn’t know.

FWIW, I personally gave up trying to make custom images, as I seem to seldom do the same thing enough times in a row that it helped more than it hurt. It’s one more kind of state to keep track of, which system images have I got & what packages are in which. Maybe if I used vscode it would be more automatic.


For the same reasons as you, I was finding the custom images not worthwhile, but as alluded to in this thread:

  • julia 1.8 makes way better custom images to begin with
  • today most packages do have useful precompile statements
  • vscode does make it much more trivial, if you have a well maintained Profile.toml and Manifest.toml (and warns you about outdated states)

Now I love custom sysimages.

And still more work is being done for 1.9 and 1.10 in order to make compiled code cache better even without custom sysimages (look on Julia’s github for pkgimages and the work by Tim Holy and Valentin Churavy)


I gave up for years because you needed to have a custom script and it would only be good for the things in the custom script. But indeed, my point is that I’ve come back to it since August. With the way that packages are setup now, it’s pretty sufficient for any decently maintained package to be snooped (not all of them are yet, but it keeps getting better), and to get a pretty good system image with a snooped package only requires mentioning it by name.

So in 2019, custom system images were not a solution. But in 2023, I think we should really be revisiting system images as something as simple as just taking the default compilation on StaticArrays + LoopVectorization + Plots can cut out >90% of the compile and load time for some of the workflows I have. Since those packages are essentially static (pun intended), I basically keep alive a “set it and forget it” system image nowadays.

VS Code’s extra piece makes this pretty trivial. Honestly the only thing that I think we’re missing is some functionality that tells you when to update the system image (because one of the packages in the system image updated), with a little popup like how updating the extension looks. That and the fact that system image building can be way faster if Faster incremental sysimg rebuilds by Keno · Pull Request #40414 · JuliaLang/julia · GitHub gets finished. Once that is finished and merged, it wouldn’t be unreasonable for a Plots-only system image update to take around 1 minute, in which case I would really be questioning why if it’s not used.

And this might make system images a thing of the past, but :person_shrugging: that’s only better.


When a bunch of stuff gets recompiled, e.g. when you install a package and frequently additional packages get (upgraded or installed or) downgraded (which you can prevent with add -preserve=all), then you can stop it with CTRL-C.

Yes, you are just deferring compilation, but often you get away with it indefinitely. At least I’m recompiling packages I might never use. The problem, as I see it, with Julia compilation is that it’s done too eagerly, and optimizing too much.

I proposed -O0 to be the default for Julia (it wasn’t wanted, I waned to opt into more compilation selectively; another option, recompile to optimize more at runtime). Julia blurs the line between development/prototyping and production, and I believe most are using the default options that are (only) better for production, where you want fastest runtime speed. For development you want fast compile times, and no inlining (–inline=no likely implied by -O0, though I’m not sure).

If your problem is “time to first plot” (not to X") then different packages have very good startup, and now this one claims “Fastest time-to-first-plot in Julia!”:

4 dependencies successfully precompiled in 3 seconds. 352 already precompiled. 2 skipped during auto due to previous errors.

julia> @time using PlotlyLight
  0.511248 seconds (491.38 k allocations: 27.544 MiB, 9.57% gc time, 79.77% compilation time)

No, you have it backwards, Apple’s ARM is king of single-threaded (and sort of recently IBM’s mainframe claimed fastest single-thread, before Apple did), last time I checked, and AMD for multithreaded.

You will have more problems in the short run with Apple’s ARM (people still run into trouble with code not compiled for it), with it a tier 2 platform (though with Rosetta it’s tier 1, and might be good enough, but then not as fast).

GPUs are irrelevant to compilation, but precomiling can eat up memory with many julia processes forked. Each one is single-threaded, but with many used at one, many “threads” are used, also multiplying memory by N. People have run out of memory in (2 GB) containers, the issue was supposed to be slowed, but seems it isn’t. There’s a workaround, you can force less or no parallel precompilation.

Ideally you would only precompile packages (or the methods of them, such as exported functions) you are likely to use in the (near) future, e.g. just the package you were just installing (and maybe some of its dependencies), often something unrelated gets changed and precompiled too.

Thank you for this! I did not know :slight_smile:

For Plots.jl or Makie.jl or other (non-plot code) this might help:

to enable PyPlot showing properly after it was baked into the [system] image add

I was looking for another project which already has some custom system images pre-built for you (e.g. for Makie.jl if I recall and JuliaSyntax.jl). It wasn’t yet announced last time I know, just be aware that there’s a third project, but that top one seems similar.

I consider myself reasonably experienced with julia, but never used system images. Using them doesn’t look as simple as sometimes portrayed, and I just don’t see any obvious little-maintenance way to incorporate them. For example, I do most of my interactive work in pluto notebooks, and would be glad to cut notebook startup time with some compilation. Are there any tutorials for this presumably common usecase?


Not about Pluto, but the way to do it should be mostly the same: GitHub - ufechner7/Plotting: Demo for using InspectDR

Each Pluto notebook has its own Project.toml and Manifest.toml file. I am not sure a sysimage would be a particularly pleasant experience in such a case. You really want the sysimage you are running to be matched to the Manifest of the project.

1 Like

Even bigger issue is that Pluto doesn’t work with custom system images last time I checked. So yes, Pluto is currently left out of the solutions right now.

1 Like

I think one problem is that using system images goes right against the principle to use a new Project.toml for each separate project and not rely on one (or a few) global environments. Also, it doesn’t help at all when developing a package, which I do most of the time. For me, what would help the most is native code caching that stays intact as much as possible when making changes to the Project.toml. I would already be using sysimages if my work was “static” enough for that.


It does. You don’t need to system image the package you’re developing. Plots, StaticArrays, FillArrays, etc. A few basic pieces chops out a ton of the compile time.

1 Like