Monday, November 10, 2008

RubyConf 2008 - Future of RubyVM - Koichi Sasada

The first session I attended at RubyConf 2008 was called "Future of RubyVM", and it began my ascent into "VM Wonkiness" as I started to call it. Koichi-san is responsible for leading the team of Ruby 1.9 implementers, and is the authoritative source of info on its status. I was hoping to learn a few things, as well as to run into new acquaintance Nobuyoshi Nakada, one of the core team members.

Techniques for VM Performance
- Simple techniques
- C level VM

- Advanced techniques
- Dynamic code generation
- Native compilation
- JIT compilation
- Polymorphic inline cache
- Selective inlining

- Online feedback opt.
- Hotspot JIT
- Tracing JIT

Pros/Cons of JRuby/IronRuby
- Using an awesome VM
- Pros
- Many clever people working on the project
- No code is good code
- Many libraries already exist on each environment
- Easy to use parallelism
- Cons
- Not only focused on ruby, so there is a semantics gap.
- Cannot use C ext direct

Pros/Cons of Rubinius
- Most code in ruby
- Pros
- Ruby in Ruby
- Says this is the *best way* to improve performance in the long run because we can easily analyze and make specific performance improvements.
- Cons
- Still a LONG way to get a high performance VM

Pros/Cons of C Ruby
- Pros
- Portability, using GCC
- C is well known language already
- Extensibility
- Performance improvement
- Easy to write simple extensions.
Cons of C Ruby
- Extension libraries are written in C
- GC problem
- Inlining problem
- Limitations on program analysis

Our Performance Policy
- CRuby is not best solution, but it is a GOOD one
- They continue to improve the CRuby implementation
- It is a pragmatic, practical selection
- It will be the main implementation still for at least several years

Keywords for Success
- Embedding
- Parallelism

The team is working on taking advantage of C some projects are running
- Hidden optimization techniques on YARV
- Ricsin

Hidden Optimization Techniques
- turned off in 1.9.1 by default
- Tail call optimization
- Optimizing using unification
- Stack caching

- Left easy optimizations
- Efficient method caching
- Efficient fiber implementation using platform dependent way suvh as makcontext()
- These optimizations will be merged in 1.9.2

Ricsin; Mix-in C Ruby
- Embed part of a C program into ruby
- Like a RubyInline, but directly embedded
- Usage
- Use C libs direct
- Replace all built-in classes and methods
- Test Ruby C API
- Continuous performance improvements

Ruby to C AOT Compiler
- Translate Ruby scripts to C code ahead of time

Related work
- ruby2c
- yajit
- yarv2llvm

Atomic-Ruby project
- Issue: ruby is too fat
- Embedded system can have special problems, such as resource limitations
- Application embedded ruby
- Some applications need just a Ruby DSL engine, not the full ruby distribution.
- We need a slim Ruby interpreter
- Utilize CRuby portability
- 3 subprojects
- Plugin/out
- Precompilation
- Swift core features

Auto-write barrier detection
- write barrier needed for several GC algorithms.
- automatic WB detection system

Generational GC
1 but ref count

Multi-VM (MVM) Project
- Multi VM in 1 process
- Watch runs in parallel
- High speed inter VM communication
- Sponsored by Sun

- CRuby/YARV is not "best solution", but it is best for right now. But eventually, Rubinius is the future. Which just so happened to be the next session I attended...

No comments: