Introducing Surinx (Ruby syntax w/ Java typing)

http://blog.headius.com/2009/08/introducing-surinx.html

by fogus • 17 years ago

51 32 17 years ago

blog.headius.com

apgwoz5• 17 years ago

> ... fib(40) in only 7 seconds on current MLVM (OpenJDK7 + invokedynamic and other tidbits), a five-fold improvement over JRuby's fastest mode (jruby --fast).

I was going to say, "damn, that's it? that's pretty slow" until I realized:

   $ time python fib.py 40
   102334155

   real    1m43.363s
   user    1m36.824s
   sys     0m0.577s

It may be the fact that I have a ton of stuff running on my machine right now, but still.

lliiffee4• 17 years ago

Does fib.py use psyco? If not, that does still seem slow. Standard numerical benchmarks indicate that c or fortran or whatever should have a much larger speed advantage relative to regular python than 40/163. E.g. on the n-body benchmark here:

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

Python is about 65 times slower than c/c++/java/fortran/scala.

apgwoz5• 17 years ago

No, it doesn't use psyco. If it did, I'd expect it to be faster of course.

I just rewrote it in C though:

   $ time ./fib 40 
   102334155
   real    0m3.063s
   user    0m2.912s
   sys     0m0.014s

Which makes the python version something like 35x slower in wall clock time.

mey• 17 years ago

Or you could just use Groovy.

raganwald5• 17 years ago

The Fibonacci implementation here:

http://github.com/raganwald/homoiconic/blob/master/2008-12-1...

Does a pretty quick job using plain-old-old-ruby (POOR Implementation):

    time ruby /Users/raganwald/Desktop/matrix_fib.rb   102334155

    real    0m0.012s
    user    0m0.007s
    sys     0m0.004s

That doesn't mean we shouldn't welcome implementation improvements, but in my lifetime the quest for a "sufficiently smart compiler" that will do away with thinking programmers has enjoyed only modest success.

Append: This is not a criticism of Charles' amazing work dragging Ruby into the 1980s and designing Ruby-like languages for the JVM. If asked to choose between algorithm improvements and implementation improvements, I answer "Yes!"

lliiffee4• 17 years ago

That basically uses dynamic programming, which is "cheating". (It is a linear time algorithm instead of exponential!) Different compilers will almost always just give us constant factor improvements. Smarter algorithms will always beat that, but sometimes a smart algorithm with a slow compiler still isn't fast enough.

raganwald5• 17 years ago

I don't think the matrix algorithm is linear, in other words I don't think fib(40) takes exactly twice as long to compute as fib(20). I think it's faster than linear, because it converts the multiplications to exponentiation and then takes advantage of composeability such that fib(40) should require on the order of one more operation than fib(20).

I think that means it's order log 2 n, but perhaps you HN folks can correct my poor math skills.

wingo2• 17 years ago

Neat hack, but needless, and besides that, irrelevant: the author didn't really care about the fibonacci sequence. He cared about function calls and additions.

See Performance and Evaluation of Lisp Systems, http://www.dreamsongs.com/Files/Timrep.pdf, for more discussions of the nature of benchmarks.

raganwald5• 17 years ago

Thank you, I am aware that the author was deliberately picking a "most pessimum" algorithm as a benchmark. I am not commenting on the nature of benchmarks, but rather on the relative importance of algorithms and implementations.

You may find it irrelevant to the author's point, others may find it an interesting byway to follow in a larger conversation. Or not.

lucifer• 17 years ago

There are so many (new) languages for the JVM; its becoming a curse of plenty! (Not complaining.) Its a great platform, and invokedynamic will usher in a whole new generation of languages.

Lets just hope the Java 7 actually happens given the change of ownership, and I would love to see John Rose's call for programmatic access from Java (http://blogs.sun.com/jrose/entry/jsr_292_support_in_javac) realized. (ASM is great stuff, but I'll admit that its not a great joy to use for me.)

donw• 17 years ago

I never stop being impressed by Charles (the JRuby lead dev)... remind me to ship a case of the Sierra Nevada Hefeweitzen to Engine Yard.

The best thing is that this will likely percolate into JRuby, and it bodes well for JRuby performance on JVMs that support invokeDynamic (not sure if OpenJDK does yet...)

rogerthat• 17 years ago

You are reminded to ship a case of the Sierra Nevada Hefeweitzen to Engine Yard.

jrockway• 17 years ago

Java has a type system?

megaduck2• 17 years ago

One of the beautiful things about JRuby is that for performance-intensive code, you can drop to Java instead of dropping to C. Java's almost as fast, and a whole heck of a lot nicer to program in.

If Surinx is able to produce Java-like speeds, that means that you never have to drop out to a different language at all. That's pretty awesome, and it would open up a whole new world to Ruby programmers.

I hope some of this work gets pushed into JRuby. I'll take 5x speed boosts any day, constant factor or not.

megaduck2• 17 years ago

One thing to remember about a lot of JVM language benchmarks is that they include the startup time for the JVM. Java code running on a pre-warmed JVM is significantly faster, and for short programs can be faster by an order of magnitude (or two).

For a good example, check out Nailgun (http://martiansoftware.com/nailgun/background.html).

snprbob86• 17 years ago

Boo is a similar project with Python and .NET:

http://boo.codehaus.org/

Introducing Surinx (Ruby syntax w/ Java typing)

Related Stories