So Google has launched the unladen swallow project with this first goal:
Produce a version of Python at least 5x faster than CPython.
We discussed some details with Collin Winter, Jeffrey Yasskin and Thomas Wouters during the VM summit yesterday. We were a bit confused about usage of the term JIT, because as far as we understood, it's going to be upfront compilation into LLVM. In the past we have looked into LLVM – at one point PyPy extensively use it but it wasn't clear how we could make good use to it. They also consider changing to something else than LLVM. It's gonna be interesting to see how this works out.
It's good to see friendly competition, and we think that should take up the challenge and see if we can produce faster pickling, run 2to3 and Django faster than what they can come up with. We also talked to IronPython and Jython developers and all agreed that some common benchmarks would be good. And maybe do weekly press releases about small speed increases? :)
The idea of the VM summit here in Chicago was to bring together implementors of various virtual machine languages. There were members of the communities of IronPython, CPython, GemStone's MagLev, Rubinius, Mozilla's TraceMonkey, Parrot, Sun's Da Vinci Machine, Microsoft's DLR, Jython and JRuby. Everybody got to talk 5-10 minutes on their current status and challenges. It is clear that you cannot begin to cover the complexities and architectures of the involved projects. But that wasn't too much of a problem because the rest of the day everybody freely and dynamically grouped on their issues of choice. We established some more personal contacts, was great to chat with people like Andreas Gal from the University of California, Irvine, who have a very similar idea about the JIT that we have. Actually, we could probably haved mixed our two presentations and nobody would have actually noticed :-).
At the end of the presentation part, John Rose presented his slides. John is a Hotspot developer, and while not precisely a dynamic language implementor, he has a lot of experience in virtual machine implementation. It's very good to see the JVM being extended towards supporting dynamic-language specific things, in order to be something more than just a good platform for Java. We'll probably have some extra meetup with him the next days.
cheers,holger and fijal
So Google has launched the unladen swallow project with this first goal
ReplyDeleteI'm not sure this is a Google project. It's hosted on Google code for sure, but anyone can do that.
All three of the primary developers are Google employees.
ReplyDeleteWe were a bit confused about usage of the term JIT, because as far as we understood, it's going to be upfront compilation into LLVM.
ReplyDeleteThe LLVM supports JIT, so compiling Python into LLVM bytecode will give JIT for free.
Anonymous#1, this is extremely valuable note to take in an open source world, you know ;-)
ReplyDeletePyPy folks, keep up the good work!
But... Sometimes I miss updates on this blog. Not in the sense that you slack on it, but in the sense that I miss some "technicaly intermediate" updates when there are no news on breathrouths.
One thing I miss most is the retro style articles on how some things that are "established" now got to be this way. The design by evolution things. Stuff that both educates and helps to get acquainted with the code one might like to hack one day.
I don't understand: These days, Google's v8 claims to be 56x faster than common javascript, tracemonkey is in the same league, as well as nitro, etc. Way before, psyco sped up python (theoretically) up to c's speed for algorithmic code, and up to 4x for common code.
ReplyDeleteNow Unladen Swallow aims to "only" 5x speed up. Isn't it to little, seeing what the above projects are getting nowadays?
Or am I getting confussed by their terminology? (what's exactly the meaning of 5x here?).
This comment has been removed by the author.
ReplyDeleteThe exact meaning of 5x is I *think* "5x on tests derived from google internal apps". It's a bit little, but note that the great speedups of JS engines are for simple algorithmic code (the one that psyco speedups great).
ReplyDeleteIt would be a good speedup for a lot of people though (of course we aim to speed up stuff according to JS engines ;)
Cheers,
fijal
Maciej, this is a reply posted on the project's FAQ page:
ReplyDeleteComment by collinw, Today (8 hours ago)
luismgz: translating Python to Javascript would be easy to implement for about 80% of the language, but you'd hit a wall in implementing that last 20%. Just ask the Jython, PyPy? and IronPython? teams how hard 100% compatibility is. They've done some really heroic work to implement every dark and musty corner of the language, and I think they'd be the first to tell you that it's easy to get something like the Fibonacci function working, but things like metaclasses are a different story. We hope to side-step that by reusing as much of CPython as possible.
Psyco's claimed benefits of 100x speed-up on algorithmic code is rarely seen in real applications. It can certainly be used to optimize hotspots that fit Psyco's profile, but in examining the performance of some Google applications that use Psyco, we found that they see only a ~10% improvement in overall CPU usage. While that might be a valuable savings for some applications, it's not 100x, nor even the 2-4x low-end estimate that I've seen in the Psyco docs.
Are our performance goals too modest? We don't think so. Our team is small -- only two full-time engineers -- and we want to allow for unexpected surprises along the way. We feel that 5x is a good goal for the time being, especially given that we may need to make changes to LLVM along the way. If things go astoundingly well, we may raise those numbers, but for now, we're comfortable with our stated goals.
Maciej said: "It would be a good speedup for a lot of people though (of course we aim to speed up stuff according to JS engines ;)"
ReplyDeleteWhat ever happend to the "secret goal" of being faster than c...?