News

Rubinius adds Multi-VM support

Posted by Werner Schuster on Jan 16, 2008

Community: Ruby
Topics: Performance & Scalability
Tags: Rubinius ,; Concurrency ,; Multi-threading

Launching multiple Ruby instances - without having to resort to executing the Ruby binary is available in a few shapes and forms. JRuby allows to have multiple JRuby instances inside one process, each running on its own native thread. For MRI, why the lucky stiff provides the Sandbox Hack to do the same thing.

Recently Rubinius got proper support for launching Multiple Ruby Runtimes, which can be used like this::

vm = Rubinius::VM.spawn "blah", "-e", "puts 'hello\n'"
vm.join
p vm.stdout.gets

To understand the implications of this, we talked to Evan Phoenix of the Rubinius project.

Asked about the idea behind Multi-VM, Evan answers:

I've always know we'd have this kind of functionality in one form or another. The plan for the feature is mixed. Yes, it makes it much easier to start a new VM to do some work. The real question is what that work is and how it's performed [..].

Evan continues with some implementation details:

Currently, each VM is started in it's own native thread (pthread on unix). This lets each VM run without any knowledge of the other VMs and keeps scheduling exactly the same inside the VM.

Rubinius is a very well behaved C program in that it uses no globals or anything, allowing multiple copies of it to mingle in the same address space peacefully.
Since they are running one per native thread, one can crash and bring the whole process down, yes.

Note that the last statement is true of all other Multi VM solutions as well - a JRuby instance that calls global, static methods like System.exit() or uses JNI code that crashes the JVM will bring down all other JRuby instances as well.

Another interesting topic is how to allow the VMs to communicate:

Pipes are one [way], and yes, stdio is redirected to pipes for sub VMs. I've also added a very simple shared message passing system. One VM can send message to another using a single, toplevel mechanism. This mechanism is the only place where one VM interacts with another. It also marshals the messages into it's shared storage, so there is no pointer sharing between the VMs.
The existing shared message passing mechanism could also be extended to use shared memory to allow messages to be passed between [OS] processes easily.

This feature is available in Rubinius git repository now (see InfoQ's coverage of using git to access the Rubinius repository) . If you just want a quick look at the Rubinius source, there's a web interface to the Rubinius git repository. For instance, here the commit for the mentioned message passing system.