2022-07-28 –, Purple
The tension between performance and flexibility is always present when developing new systems. Often, poor performance is unacceptable. But poor flexibility hinders experimentation and evolution, which may lead to bad performance later on. In this talk, we show how we used MMTk.io – a toolkit we are developing that provides language implementers with a powerful garbage collection framework – to implement a flexible (unbaking the cake) and performant (and eating it too) memory manager for Julia.
When implementing a system such as a programming language runtime, decisions usually favor performance over flexibility. Lacking performance is often unacceptable but lacking flexibility can hinder experimentation and evolution, which may also affect performance in the long run. Consider memory management, for example. If we ignore flexibility and only favor performance, sticking to a particular type of garbage collector that "performs well" can have a huge effect later on, such that changing any aspect about it can be almost impossible without rewriting the whole system.
MMTk.io is a memory management toolkit providing language implementers with a powerful memory management framework and researchers with a multi-runtime platform for memory management research. Instead of a single, monolithic collector, MMTk efficiently implements various garbage collector strategies, increasing flexibility without compromising performance.
MMTk started with its original Java implementation, which was integrated into Jikes RVM in 2002. Since then, it has gained a fresh Rust implementation, which is under active development and even though it is not ready for production use, can currently be used experimentally.
To use MMTk, one must develop a binding, which contains three artefacts: (i) a logical extension of the VM, (ii) a logical extension of MMTk core, and (iii) an implementation of MMTk's API. At the moment, there are various bindings under development including bindings for V8, OpenJDK, Jikes RVM, Ruby, GHC, PyPy and now Julia.
In this talk we discuss our experience developing the MMTk binding for Julia. Julia currently implements a precise non-moving generational garbage collector. It relies on some LLVM features to calculate roots, but the code follows a monolithic approach, as described earlier.
We reuse some of Julia's strategies for calculating roots and processing objects, integrating these into an Immix implementation inside MMTk. Our implementation passes all but a few of Julia's correctness tests, and has shown promising results regarding GC performance. We hope that with MMTk-Julia we are able to easily explore different GC strategies, including a partially-moving GC, observing how these strategies affect the language's performance.
With a broad interest in programming language design and implementation, I'm currently a PostDoc at the Australian National University working on memory management and virtual machines. My previous work from my PhD at Delft University of Technology focused on syntax definition formalisms and parsing.