Clojure 1.3 First Impression (It's Fast) | Hacker News Reader

Clojure 1.3 First Impression (It's Fast)

learningclojure.com

102 pointsjohnaspden25 comments

Last time I tried Scala, it blew Clojure out of the water for speed on the program I translated. I assume that Java did too. I think that's probably not true any more.

Anyway, pretty fractal tree program in lisp!

You're still going to find Scala to be significantly faster on arbitrary code. It's great that there are ways to speed up key functions in Clojure now though.

Backed up by what evidence? And by arbitrary code do you mean code that embraces mutability?

Scala effectively has these kinds of type declarations on every function. This article itself proves the point. Take out the :static and type decls and see.

Idiomatic Clojure tends to use Clojure's persistent data structures - you don't often create custom types, all of the core functions on those data structures are already pretty much as fast as possible. No need for ^:static or type decls.

The article also doesn't get into deftype/defrecord/protocols. In those cases you always get the fastest path of the platform. Again you can build things that have the same perf of the core data structures w/o resorting to ^:static or type decls.

So what exactly do you mean again by arbitrary code? Perhaps you meant numeric code - there ^:static and type decls help plenty. Perhaps you mean Java interop? Again sure.

Personally I think Clojure is really taking the dynamic, generic, and fast thing to a whole new level, w/o type-hinting of any kind.

EDIT: This isn't too say Clojure performance can't continue to improve. Scala's ability to use Java arrays of primitives in higher order operations is something I really, really want to see (and the the above is a big chunk of the work in that direction)

It's strange that you mention Clojure's persistent data structures as an advantage for performance. They have many advantages but performance isn't one of them.

I'm curious how you come to this conclusion? Programs in any language will often be designed around immutability in order to improve code maintainability. Without fast persistent data structures this will most certainly decrease the performance of your otherwise elegant design due to considerable amounts of copying.

It's threads like this that make the Clojure community really irk me sometimes. It's ok to think your language is great, but please take a reality check before you go off and try to defend it against criticism! Judging by your comments, especially this fuzzy little one about elegant design, I would guess that you've spent very little time dealing with actual CPU and memory-related performance issues.

It's not just me then.

If you're going for raw speed you should use mutable data structures. Sadly, that's just the way it is - you can update a mutable data structure in a couple of instructions by swapping out a pointer, while "updating" an immutable/persistent object is inherently more work at the machine level.

Clojure is awesome because it provides immutable persistent structures that are well onto the good side of "good enough", not because it's the fastest thing possible. Java itself is far from the fastest thing possible, and the speed of Java is itself a strict upper bound on Clojure's speed.

Sure, but of course if you want raw speed you would write assembly. Most people move up the food chain to C to get some basic level of correctness.

I think the history of production oriented PLs has revolved around balancing correctness/performance. And it's always a question of diminishing return (I've written enough C/C++/Objective-C++ to know this). Every sizeable application I've written in a lower level language eventually approaches some kind of wacky manual memory management scheme to preserve performance w/o sacrificing correctness. I think Clojure's solution is a step in the right direction.

But no, if I was writing a real-time system with extremely limited memory constraints I would not write in the Clojure. But I don't think that's what this thread is about.

There is a large class of problems that aren't worth writing in assembly but for which Clojure is (currently) still too slow.

Sometimes you just need something at Java's (or preferably Scala's) position on the power/performance spectrum.

That said, the way Clojure is improving, I think it will continue to asymptotically approach Java's speed anyway.

Being a sometime embedded man myself, I was nodding sagely at your comment until this bit: "Java itself is far from the fastest thing possible, and the speed of Java is itself a strict upper bound on Clojure's speed"

Clojure doesn't compile to Java, it compiles to JVM bytecode, and I think (although I don't know) that that's supposed to be very fast.

In this post:

http://www.learningclojure.com/2010/09/clojure-faster-than-m...

I compiled a (lookup into a map (potentially constructed at run time)) into a series of nested if statements, which then ran very fast indeed.

I shouldn't think I've achieved anything like the holy grail here, it all goes wrong at the end anyway, and I think that clojure 1.3 might actually have broken the optimization techniques I used.

But I think there's a case for saying that an immutable lisp on the JVM might one day be the fastest language in the world, because you can do those sorts of tricks in lisp in a way that you can't feasibly do in assembler, and the JVM does optimization dependent on run-time behaviour, as well as on static analysis.

I don't have the data. I'm not asserting this.

But I am asserting that Java speed is not a strict upper bound on JVM language speed, (unless every JVM bytecode program is also a Java program, which I can't imagine is the case!).

By the way, compiling a data structure to a lookup function is a very clever trick, and fortunately something Clojure makes easy. It reminds me of how old-school assembly programmers would blur the lines between code and data for better performance. Only this is more structured and easier to manage.

There are a couple of drawbacks, though:

1. It really only works for read-only data. Generating a new lookup-fn is a pretty expensive operation. I don't think I could use this for data that changed at all.

2. Compiled code uses permgen space and isn't garbage collected as easily. If you make extensive use of this technique, especially with large maps, it'd be really easy to blow your permgen space.

You're right, of course. I should have said JVM instead of Java.

That said, in reality, JVM bytecode is tied pretty closely with Java. The bulk of a program: method calls, types, loops, etc all map pretty much 1:1 with what Java emits - there's not a whole lot of room for improvement. And if there were, it would probably be through MORE static typing and compile-time analysis, not less.

I expect that for any arbitrary Clojure program you give me, I can write a Java program that will execute as fast or faster. Of course, it might take me ten times as long to write and have ten times the amount of source code.

Of corse scala is fast with stuff like that because its typed. But should we talk about all the stuff you can't do because you have types all the time?

I think clojure is getting really close to a dynamic language that can get down to the speed of a static language if you need it.

Yes, let's. You first!

And Ocaml is faster than Ruby!

I was pointing out that the comparison between Scala and Clojure is between a static language and a dynamic one, the static one will almost always perform better.

As someone familiar with Java/Scala but not Clojure, what exactly do those type hints do? Just prevent it from autoboxing the primitives? Avoid excessive reflection? Can this be done for other types or just int/long/float/etc?

type-hinting Java objects to avoid reflection has been possible for some time now. What wasn't possible was for fns to take or return primitives w/o boxing. ^:static lets you do that. ^:static also allows the JVM to apply the most aggressive optimizations to your code - ^:static tells the JVM the code won't change - so callers of ^:static fns won't get the latest version if you redef them. The benefit being that with numeric code you'll often see an order of magnitude performance jump.

So a general strategy is to write your code as you normally would - def'ing redef'ing fns at will. Then when it works - declare the critical paths ^:static, and adding primitive type hints ^long, ^double if the code is numeric.

CMUCL, dealing with a similar infinitely-dynamic problem in Lisp, had a similar approach but dealt with it on a block-of-code level instead of per-function. You could declare a block to be compiled together, and any calls between functions in the block would be treated as having a promise that you'd never redef them without recompiling the whole block, so they could be statically optimized with respect to each other.

Not sure if that's a better or worse approach overall. It probably depends on the structure of your program. One nice feature was that the same function could be static from some perspectives and not from others--- to functions in the same compilation block it was static, but functions from outside that block that called it treated it as non-static, and would immediately get the new version if, say, that whole block were recompiled. That let you make the core code optimized by sticking it in one big compilation block, while not changing the normal dynamic Lisp semantics of the block when viewed from the perspective of any outside function.

Clojure supports the static from some perspectives and not from others pattern, you just have to go through the var. So even if an fn foo is declared ^:static, you can call #'foo instead so that you always get the latest version of the fn.

The diffrence form Java/Scala is that Clojure dosn't type check (like say typed scheme). Is just for performance.

typhinting to avoid reflection can be done for all java objects. The autoboxing is only for numeric stuff.

It avoids reflection, and instanceof checks.

Reply on news.ycombinator.com