Love for LuaJIT | Hacker News Reader

Love for LuaJIT

google-opensource.blogspot.com

75 pointswhyenot45 comments

LuaJIT 2.0 is an amazing achievement. On the programming language shootout's "flawed benchmarks", LuaJIT outperforms almost every other dynamic language (including JavaScript V8, Python, Ruby 1.9, JRuby, VisualWorks Smalltalk, CPython, PLT Scheme), and is competetive on many benchmarks with highly-optimized language implementations like Sun's JVM, and GHC:

http://shootout.alioth.debian.org/u32q/which-programming-lan...

Mike Pall wrote LuaJIT and now LuaJIT 2 as essentially solo efforts. Thanks to Google and the other sponsors for helping him continue!

For an apples-to-apples comparison you'll want to look at the single-threaded performance:

http://shootout.alioth.debian.org/u32/benchmark.php?test=all...

Spoiler: LuaJit compares even better when comparing single-threaded performance.

Edit: replaced "performs" with "compares", eliminating ambiguity around competitive performance vs clocktime performance

Spoiler: The Lua benchmark submissions are not parallelized.

In my experience, the 1.x releases of LuaJIT have typically made whole programs (not just benchmarks) run 2-3+x faster than standard Lua. Without any other changes. :)

I haven't tried LuaJIT 2 yet, but I'm eagerly waiting for it to get ported to amd64.

The fact that luaJIT is written in assembler makes me wonder how practical an amd64 port. :-\

Most of the VM is written in C. Only the core of the interpreter is written in assembler (generates ~13K of machine code).

And the port of the interpreter to x64 is already done, so it was certainly practical. Porting the rest of the VM and the JIT compiler is what takes most of the time.

Thanks for the clarification!

it would be interesting to see LuaJIT as part of LuaTex.

Super fast desktop publishing FTW !

I was abusing a footnote package, sorting footnotes alphabetically per page, and that took about a second a page on my laptop, but for everything else disk seeks take the majority of TeX's time; once you LaTeX something, and cache the font files in memory, it usually takes under a second or two for anything under 50 pages.

We use Lua internally at Google

Interesting! Anyone know details?

Joshua Haberman gives some details here: http://article.gmane.org/gmane.comp.lang.lua.general/62321

Our Lua usage isn't too widespread at the moment; it's really one infrastructure project in particular that uses Lua to allow user-defined functions to run within a tightly controlled container. Lua was the best choice, because of its low overhead, fast execution, and the ability to set limits on execution time.

Unfortunately I cannot be more specific, since the project is not public.

Man, you only have to say something once and then the Internet replicates it for you. :)

Well, if there is someone motivated enough to search for it.

and a company with a mandate to publicly index everything...

...well, almost everything.

Interesting to read this. I thought Google only used C and JVM-based languages as a company policy. Perhaps that's still the case except for this one example?

Just a guess google power meter? http://www.google.org/powermeter/index.html

Mike Pall's annoucement on Lua list: http://thread.gmane.org/gmane.comp.lang.lua.general/62316

It's been on my todo list to take some time to grok luajit's internals for awhile. Awesome to see Google parting ways with some cash!

This sort of makes me want to try to write a Python->Lua compiler.

Anyone using Lua for web apps? What's the current state ofthe ecosystem? Anything good, stable, and easy to deploy?

That's something I've also been wondering lately, since the Kepler and Orbit projects haven't seen activity for ages (at least according to their webpages).

I found this Sinatra clone for Lua called Mercury (http://github.com/nrk/mercury), but it lacks documentation and does not seem to be very mature at the moment.

Yeah, I've poked around a bit as well, but didn't find anything that seemed like it was getting traction.

I was excited when I heard that mod_lua was going to be included with Apache by default [1] ... but I haven't really heard anything since then.

[1] http://news.ycombinator.com/item?id=452396

LuaForge's ftp server is broken since the big LuaForge crash a few months ago, so those pages are badly out of date. Browse http://github.com/keplerproject for more recent information.

it makes me wonder why languages like ruby are so slow.

Ruby might just need a better implementation with more time spent on optimization. http://maglev.gemstone.com/ is going in this direction.

But also, Lua was always designed with cleanness and simplicity of implementation in mind. Its developers have strong backgrounds in both theory and engineering, which you can see in their interview in the book Masterminds of Programming. They've been very conservative about the design of the language. In contrast, Ruby feels much messier and more "evolved" - look at its Yacc grammar[1], for example, compared to the grammar of JavaScript or Lua[2]. Even the Ruby lexer is full of weird edge cases. Much of MRI is like that, and for years it was the only "spec" for the language; now its decisions have been reverse-engineered and inherited by all other implementations.

[1]: http://blog.nicksieger.com/articles/2006/10/27/visualization...

[2]: http://www.lua.org/manual/5.1/manual.html#8

[By the way, I've used Ruby for years and like it at lot, while I've written only a few lines of Lua code and like it but am not deeply familiar with it.]

Also, Ruby has much more complicated rules for method dispatch. Lua just has functions, and if you want to implement any more complicated dispatch, you have to do it yourself. Once I implemented Ruby-like dispatch on top of Lua, and it was terribly slow:

http://blog.reverberate.org/2009/02/09/disappointment/

To make a long story short, porting my existing Lua project to use my Ruby-like method dispatch resulted in a 4x slowdown.

Also, "The Evolution of Lua" — an interesting paper that shows how they came to this fast and somewhat elegant language: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.72.... [PDF]

Also good: "The Implementation of Lua 5.0" (http://www.tecgraf.puc-rio.br/~lhf/ftp/doc/sblp2005.pdf).

Past discussion: http://news.ycombinator.com/item?id=617278

If you want to learn Lua, Ierusalimschy's _Programming in Lua_ (http://www.inf.puc-rio.br/~roberto/pil2/) is by far the best book. Lua is one of my favorite languages. While it's good on its own, you'll get a lot more out of it if you're comfortable with C. Lua+C is an excellent combination.

Just chiming in as another Lua fan; I really like it because it's like the Scheme of imperative scripting languages. And while we're passing around links to papers, I've found the following the be good for additional insight:

Lua Performance Tips: (http://www.lua.org/gems/sample.pdf)

A No-Frills Introduction to Lua 5.1 VM Instructions: (http://luaforge.net/docman/view.php/83/98/ANoFrillsIntroToLu...)

Agree. If you want to learn more about Lua+C check out sources of this awesome game: http://springrts.com/

Ruby might just need a better implementation with more time spent on optimization...But also, Lua was always designed with cleanness and simplicity of implementation in mind

Ease of implementation has a multiplier effect with Open development effort. Not only do you get more of "more eyeballs," new contributors can ramp up faster. There tends to be a better effort/ fun ratio.

Even the Ruby lexer is full of weird edge cases.

Actually, the lexer sets state that changes the operation of the parser and vice versa.

It'd be interesting to see how fast Ruby would run "cached" (i.e. with the AST already assembled in memory,) to see how much of the time is really spent in the parsing and lexing phases.

Note that I'm not claiming Ruby's parser and lexer are slow, or that they're the cause of any greater problems. I'm just picking on them as one example of deeper differences between the Ruby and Lua designs/implementations.

Some of Ruby's more excentric semantics are really hard to optimize away by a compiler. It takes a much greater effort to compile Ruby source code to machine code of similar quality than for Lua. Since time and money are not infinite resources, especially for compiler development, you'll get worse results in practice.

So, yes, language design does impact performance.

> Some of Ruby's more excentric semantics are really hard to optimize away by a compiler.

Any examples in particular? Besides name resolution and method dispatch (and callcc, but I think that may have even been removed from Ruby 1.9) none of Ruby's semantics seem all that different from Lua. Am I missing something?

I guess there's also throw/catch and exceptions.

Here is a step by step comparison on how cutting off ruby's features can make the JRuby implementation faster, from the guy who worked on JRuby.

http://blog.headius.com/2009/04/how-jruby-makes-ruby-fast.ht...

Read it, you will find it interesting. The guy also wrote a small language called Duby which used Ruby like syntax but is much more faster, because he didn't include any of those slowing down features.

Mostly resolution and dispatch. Some of this needs very clever caching or the overhead can't be eliminated or hoisted. The control-flow constructs and iterators come to mind, too.

The Rubyists certainly know about these issues, so you'll have to ask them for details.

There's also the over-featured thread implementation (especially before 1.9)

Dispatch is worse than you realize due to the hackish optimizations made in MRI: if you monkeypatch a natively-method from Ruby, other native code will call the original by default.

Ruby was just designed and implemented with all the attention paid to sugary syntax/semantics and very little thought given to how well it could be implemented. MRI is a really poor implementation of a (formerly) poorly-specified, highly irregular, unorthogonal language.

MRI's garbage collector is abysmal, use Phusion's REE fork to get one that at least doesn't mark every single object on every collection and actually bothers to perform them in normal situations. It didn't do any desugaring at all until 1.9, and YARV only has bytecodes for the most primitive accessors, assignments, stackops, jump/condjump, and returns; with optimizations for -, +, <. <<, and some regexes. Your normal code just gets passed through verbatim.

Every once in a while I'll be reading docs or sitting in irb and suddenly scream at the ceiling, spin in my chair, and get the urge to throttle a Japanese person.

No need to wonder, have a look at the source code for them both and it will become clear exactly why there is such a performance difference. Lua gives you raw C numeric types which makes a big difference right there. Other people have mentioned a few other good reasons too (no costly method dispatch).

I haven't looked at LuaJIT yet but considering how small it is I'm sure it meets or exceeds the quality of the code behind Lua.

> Lua gives you raw C numeric types

Standard Lua (the one from PUC-Rio) doesn't really, it uses boxed numeric values, as I imagine ruby (and almost every other dynamic language) does as well. "Boxed" meaning that each number is stored on the heap with some kind of tag, and when you pass such a value to a function you're really passing a pointer to it and that indirection costs cycles and puts pressure on the cache. LuaJIT's most significant (IMO) feature is that it stores numeric values directly, like C does, so when you call a function the actual bits representing it go right on the stack, not the address of where the bits can be found, and it uses some cool tricks of floating point representation to prevent the garbage collector from trying to follow these stack values as if they were pointers. This (combined with the trace optimization) is why when you write a numeric for-loop in LuaJIT it runs (nearly?) as fast a C.

Lua numbers are not boxed in the usual sense; they live in the Lua data stack (which leaves in the C heap, but this is not isomorphic to the Lua heap), are not accessed through a pointer, and do not put pressure on Lua's memory allocator and garbage collector.

LuaJIT 2's NaN tagging alone does not speed up the stock interpreter: http://article.gmane.org/gmane.comp.lang.lua.general/44823.

http://www.macruby.org/ -- Apple's supporting an OSS project to use LLVM for real compilation, getting Ruby within spitting distance of Objective-C.

Of course, the fact that it's tied to the Objective-C runtime means it's not portable as-is, but at least it's a proof that Ruby need not be slow.

Can't wait for the x64 version.

Reply on news.ycombinator.com