Julia pre-compiles packages on first load, allowing them to deeply optimize the generated code for the particular CPU architecture the user is running on for maximum performance. However, it does take some time, so there’s a startup latency penalty here. If you make docker images with precompiled Julia packages, you can pay this pre-compilation penalty at image build time, rather than at image startup time. Much better for your users!
If you look at the tutorial for pre-compilation, it may sound like all you need to do is call
Pkg.precompile. And if you actually build and test your docker image, it will work fine - your precompiled libraries are loaded, and startup is fast. Yay!
But then you push your built image to a registry, and someone else pulls it and runs it - and the pre-compilation has no effect! Packages are pre-compiled again on first run, and sometimes this may take multiple minutes, causing weird to debug timeouts. But then you can’t reproduce it, because it runs fine on your computer! What’s going on?
What is going on is an abstraction leak. We think of computers as ‘computers’, and often it’s easy to forget that there’s just a wide wide variety of them. After a few decades of Intel’s (or should I say AMD’s) amd64 architecture hegemony, M1 Macs (and Raspberry Pi) have at least forced us to consider that ARM64 exists as well. But those are still just broad classes of architectures, and each generation of CPUs within them also have differences, with specific instructions available in some that aren’t in others.
So turns out Julia’s pre-compilation is CPU dependent (or more specifically, ISA dependent). So if I build my docker image on one type of CPU, the pre-compiled files will only be used when running on the same type of CPU!
The limits of ‘works on my machine’ :)
This manifested as a ‘Pluto notebook is not starting when run on JupyterHub’ issue. It was fun debugging this, eventually leading me to discover this bit about Julia pre-compilation. We eventually landed a fix that cross compiles to a number of different ISAs, producing a ‘fat binary pre-compilation’ that starts fast on a number of different CPUs!
I’m also personally happy with the amount of time this took - I’m trying to learn to wean myself off high stress ‘gotta fix it now’ open source maintenance, and a slower paced but more consistent approach.
Anyway, if you’re building Docker images with Julia in them, and want to pre-compile any packages you install in the image (as you should), I recommend the following bash snippet before you call
if [ "$(uname -m)" == "x86_64" ]; then # See https://github.com/JuliaCI/julia-buildkite/blob/70bde73f6cb17d4381b62236fc2d96b1c7acbba7/utilities/build_envs.sh#L24 # for an explanation of these options export JULIA_CPU_TARGET="generic;sandybridge,-xsaveopt,clone_all;haswell,-rdrnd,base(1)" elif [ "$(uname -m)" == "aarch64" ]; then # See https://github.com/JuliaCI/julia-buildkite/blob/70bde73f6cb17d4381b62236fc2d96b1c7acbba7/utilities/build_envs.sh#L54 # for an explanation of these options export JULIA_CPU_TARGET="generic;cortex-a57;thunderx2t99;carmel" fi
This is what the official Julia binary is built with (and you should be using that, not your distro provided binary). It may need updating as new ISAs come out, but otherwise should help.
P.S. Did you know that Python also does pre-compilation, and that can also have a massive effect on python startup time? I did a similar debugging session on that a while back.