JRuby Serial Interview 3
Posted by Nick Sieger Thu, 18 Jan 2007 03:59:31 GMT
This is part 3 in our ongoing conversation tracking the development of JRuby.
Official Rails support in February? That’s not far away! What do you mean by “official”?
Thomas Enebo: Largely, we just want to spend some extra TLC on fixing up various Rails issues between now and February. Basically, get rid of the remaining known issues with running Rails from JRuby. I think beyond marshalling we are very close to saying that today. We also want to provide a better deployment picture for Rails by then. So we will need to spend some time on that as well (the community has been doing a great job spearheading this).
Software is never perfect, so we know that there will continue to be Rails issues after we say it is supported. By setting this goal, we should give ourselves some pressure to polish what we have and also get a larger number of people some incentive to kick the tires.
Ola Bini: That’s a question of interpretation. In my view, “official” is some kind of high number. 95% of all test cases in 1.1.6, maybe? But the more important part of it is that all the common use cases should work. You should be able to work through AWDwR and everything should work.
It’s ambitious, but we can do it. What’s needed soon is to decide what needs to be done with ActiveRecord-JDBC, and do that, since AR-JDBC is one of the larger points in our Rails support, and sometimes I feel that the support there is our weak link.
Charles Nutter: We could probably say we support Rails today, with a whole list of caveats. Rails runs right now, people are using JRuby on Rails today (in some cases for production apps!), and things largely will “just work”. There’s also a lot of community effort behind alternative deployment scenarios like within a WAR file or behind a fast HTTP server like Grizzly. Rails does run on JRuby today.
Our challenge before making a big official announcement about “Rails support” is to shrink that list of caveats down as small as possible. We want marshalling to work so sessions function correctly. We want AR-JDBC to be cleaned up a bit more, with more testing and even wider database support. We want any remaining core library issues resolved. We want those peripheral deployment projects to work perfectly. There’s a lot of work to get there, but it’s now simply an incremental process; Rails runs today, and will run better tomorrow.
What’s this about YARV instruction interoperability?
Readers: here’s an IRC recap for you:
[1:31pm] olabini: HAHA
[1:31pm] olabini: YEAH BABY.
[1:31pm] olabini: hehe.
[1:31pm] olabini: echo 'puts "Hello world"' > test1.rb
[1:33pm] olabini: ruby-yarv ~/src/yarv/tool/compile.rb -o test1.rbc test1.rb
[1:33pm] olabini: jruby -y test1.rbc #=> "Hello world"
[1:33pm] headius: hah, awesome
[1:33pm] nicksieger: no way!So this is awesome that you guys are able to track so closely with YARV’s progress, but why do it? Hedge your bets?
Ola Bini: First of all, it shows the maturity of the JRuby runtime that we can implement basic parts of the YARV VM this easily. Second, it can be very interesting for us to try out parts of how 1.9/2.0 will work out within the current JRuby system. Third, we don’t have a YARV compiler yet, but being able to run files compiled with YARV ensures that we stay on track for Ruby compatibility. Fourth: Yeah, hedging your bets is always a good idea. Diversity breeds evolution. I believe where on the right track with the current AOT and JIT compiler works, but there is always a good idea to implement these things in more than one way. And of course, it ‘s just fun!
The next step for this will be to get the loading to handle more things. At the moment, it only runs extremely simple scripts. But I’m planning on handling the more complex things soon too, adding support for labels and defining YARV methods and other such things. The very next step to handle is a compiled recursive fibonacci script. The YARVMachine already has most things needed for it, but the YARV emitted by the compiler contains some tricks that needs to be fixed.
Charles Nutter: I’ve been the primary person responsible for our various interpreter rewrites over the past year or two. Originally I modified it to be mostly “stackless”, using an Instruction object for each element in the AST and pushing down an external stack of previous instructions to maintain context. That seemed like a neat idea, and it certainly did move toward a stackless design (I actually demoed a recursive fib(100_000) at RubyConf 2005), but it was rather slow and complicated enough that only I could maintain it. So in October of 2006 I did another rewrite, basically simplifying the interpreter down to a “big switch statement” that could quickly traverse the AST without a lot of objects and stack manipulation. This new C-like interpreter engine was quite a bit faster, but unfortunately the tradeoff was that we were again burning Java stack frames when we had to dig deeper into the AST, and our maximum stack depth suffered.
I’ve still always wanted the stackless design back in JRuby, and started to think about alternate routes to get there. The most obvious was having our own bytecode engine. The most readily-available set of bytecodes for Ruby...was YARV’s.
Implementing a stack machine is pretty trivial. You need an operand stack, instructions for manipulating it, and instructions that consume values from it. Over the past year, YARV’s core set of bytecodes have started to solidify to the point that I figured implementing a YARV machine in JRuby would be a good idea. So I did a very simple initial implementation that just handled local variables, method calls, while loops, and so on, to see if it would be feasible. I got it as far as running an iterative fib, and the results were very promising: it was quite a bit faster than the current interpreter.
I ended up putting that down for a while to look at Java bytecode compilation, which has been coming along very nicely. Recently, Ola decided to pick up the YARV machine work, and made it double cool by loading real compiled YARV bytecodes into it. And if that wasn’t good enough, the original partial machine I’d implemented was able to run them without modification!
We’re looking toward the future with this work. We know that YARV is now “The” Ruby VM, and that eventually people will start to run compiled Ruby bytecodes. We also know that JRuby will never be able to fully escape interpreted-mode execution. I believe both goals are answered by having a fast Ruby bytecode machine that works in concert with our Ruby-to-Java compilers. And that’s the current roadmap for JRuby’s execution model.
Thomas Enebo: Personally, I would like to see us move from walking our current tree to walking a set of instructions. I think dynamic optimizations (as well as static) will be much easier using instructions and we can also create a simpler AST/parser (the traditional AST in Ruby does quite a bit of static optimization which I think makes the grammar a little tougher to wrap your head around).
Anything YARV-related sort of hits my sweet spot since it may nudge us in this direction. JRuby will probably always mix execution between compiled and interpreted code. I think interpretation will be easier to support at an instruction level. It will also give us a good opportunity to flatten the Java stack more. YARV is a good place to start. It may be right solution for us too. Who knows? I think experimentation is the name of the game for this stuff. So I love to see it happening :)
