Julia Users Berlin / Mar 18 2019
Simon DanischPackageCompiler
The Problem
- Julia is statically compiled at runtime
- Speed like C++, dynamic scripting like with Python
=> buy into the same compilation times as C++ ... EVERY TIME!
Atom or Revise.jl
The solution:
- to just keep Julia running
- Julia only needs to compile small diffs
Still problematic:
- employ to environments where Julia can't compile (wasm, embedded, ...)
- deploy software to users that require maximum usability (no Julia install, no patience for startup times)
- Wrappers in other languages for Julia packages
- running Julia scripts in the cloud (time is money and it's impossible to keep Julia running)
- developer convenience (after all, even with Atom things can crash and require restart)
The Solution: PackageCompiler
ahead of time compilation in Julia:
test1(a, b) = a + b test2(a, b) = a + b # compile only one precompile(Tuple{typeof(test1), Int, Int}) test1(1, 2) test2(1, 2);
Compiling a Package Ahead of Time
apt-get update
apt-get install libcairo2 -ypkg"up; add PackageCompiler AbstractPlotting CairoMakie MakieGallery StatsMakie ColorBrewer"
using PackageCompiler sys_new, sys_old = PackageCompiler.compile_incremental(:CairoMakie);
using CairoMakie, AbstractPlotting scatter(rand(10))
cp(sys_new, sys_old, force = true) # replace system image
"/usr/local/julia/lib/julia/sys.so"
Outlook
When working nicely, this can be bigger than the combination of Python & C/C++!
- Julia is still fully dynamic, even for the AOT compiled parts
- even more so: one still has all the cool Julia features like @which or @edit available even for that parts that would usually require C/C++
- Interprocedural Optimization is still possible in all cases!
- One can decide what to compile dynamically!