Is "Performance" Subjective or Objective in nature?

(Editor's note: This post is likely to open a huge can of whoop-*ss on this blog, so unless you want to get caught up in the huge bar fight that's about to break out, you're advised to take your whiskey or beer and head outside for a smoke until the cops come.)

As a fellow Scala writer, I've been following Daniel Spiewak's blog with no small amount of interest, as he discovers little tidbits inside the Scala language (like the Option type). Then I ran across this entry, about benchmarks and comparing the performance of Java, Groovy and Scala:

I’ve seen these results dozens of times (looking back at the post), but they never cease to startle me.  How could Groovy be that much slower than everything else?  Granted it is very much a dynamic language, compared to Java and Scala which are peers in static-land.  But still, this is a ray tracer we’re talking about!  There’s no meta-programming involved to muddle the scene, so a halfway-decent optimizer should be able to at least squeeze that gradient down to maybe 5x as slow, rather than a factor of 830.

That's a huge discrepancy, and like Daniel, I'm not sure where the perf hit comes from, particularly when we consider that JRuby, another language with equally powerful metaobject protocol (MOP) capabilities, is turning in performance times that are equal to those we see with the original Ruby interpreter (according to Daniel's blog entry, though I note that the comparison of JRuby to Java isn't given). And if the disbelievers in the crowd are starting to tune this out based on the fact that "Ah, it must be an edge case, after all, there's always one benchmark that any language will fail compared to another one; maybe Groovy's just not cut out to do ray-tracing. Yeah, that must be it. Besides, how often do I really do ray-tracing when I'm writing code at work?", take heed, for Daniel notes this and starts to cite other evidence that seem to establish a disturbing pattern:

If this were an isolated incident, I would probably just blow it off as bad benchmarking, or perhaps an odd corner case that trips badness in the Groovy runtime.  Then a week later, I read this post by Pete Knego (which shows Groovy's performance as equally disappointing, on the order of 7.6x to 56x worse than equivalent Java code --TKN).

All of this is old news, so the question is: Why am I bringing this up now?  Well, I recently saw a post on Groovy Zone by none-other-than Rick Ross, talking about this very subject.  Rick’s post was in response to two posts (here and here), discussing ways to improve Groovy code performance by obfuscating code.

Uh, oh. I don't know about y'all, but anytime somebody is suggesting improving performance by obfuscating code, I'm nervous--almost by definition, code obfuscation makes code run more slowly, not more quickly, because now the bytecode is pulled out of familiar patterns recognizable by the JITter and therefore more aggressively turned into optimized native code. I'm not saying Rick is wrong, but if his experiments are leading him to understand that obfuscated code is somehow running faster than non-obfuscated code, then something deeply strange is afoot.

(Editor's note: Better hurry and head outside folks, the Groovyists in the corner are starting to grumble amongst themselves, working up the courage to toss that first beer in the piano player's face.)

Daniel's not done here, though, and goes on:

Final result?

This text is being written as I was changing and trying things, I gained 20s from
minor changes of which I lost track. :-) I am currently at 1m30s (down from the
original 4m and comparing with Java’s 4s).

I’m sorry, this is acceptable performance?  This is someone who’s spent time trying to optimize Groovy, and by his own admission, Groovy is 23x slower than the equivalent Java code.  Certainly this is a far cry from the 830x slower in the ray tracer benchmark, but in this case it’s simple string manipulation, rather than a mathematically intensive test.

Coming back to Rick’s entry, he looks at the conclusion and has this to say about it:

Language performance is highly overrated

Much is often made of the theoretical “performance” of a language based on benchmarks and arcane tests. There have even been cases where vendors have built cheats into their products specifically so they would score well on benchmarks. In the end, runtime execution speed is not as important a factor as a lot of people would think it is if they only read about performance comparisons. Other factors such as maintainability, interoperability, developer productivity and tool and library support are all very significant, too.

Wait a minute, that sounds a lot like something else I’ve read recently!  Maybe something like this:

Is picking out the few performance weaknesses the right way to judge the
overall speed of Groovy?

To me the Groovy performance is absolutely sufficient because of the
easy integration with Java. If something’s too slow, I do it in Java.
And Java compared to Python is in most cases much faster.

I appreciate the efforts of the Groovy team to improve the performance,
but if they wouldn’t, this would be no real problem to me. Groovy is the
grooviest language with a development team always having the simplicity
and elegance of the language usage in mind - and that counts to me.  :-)

This is almost a mantra for the Groovy proponents: performance is irrelevant.  What’s worse, is that the few times where they’ve been pinned down on a particular performance issue that’s obviously a problem, the response seems to be along the lines of: this test doesn’t really show anything, since micro-benchmarks are useless.

I’m sorry, but that’s a cop-out.  Face up to it, Groovy’s performance is terrible.  Anyone who claims otherwise is simply not looking at the evidence.  Oh, and if you’re going to claim that this is just a function of shoe-horning a dynamic language onto the JVM, check out a direct comparison between JRuby and and Groovy.  Groovy comes out ahead in only four of the tests.

Uh, oh.

(Editor's note: Head for the doors, folks--those guys in the corner wearing the black leather jackets sporting the "Grails Rulez" logos on the back have started to head for the center of the room, and they're looking drunk, mean, and angry.)

Here comes the coup de grace

What really bothers me about the Groovy performance debates is that most “Groovyists” seem to believe that performance is in the eye of the beholder.  The thought is that it’s all just a subjective issue and so should be discounted almost completely from the language selection process.  People who say this have obviously forgotten what it means to try to write a scalable non-trivial application which performs decently under load.  When you start getting hundreds of thousands of hits an hour, you’ll be willing to sell your soul for every last millisecond.

The only answer I can think of is that the Groovy core team just doesn’t value performance.  Why else would they consistently bury their heads in the sand, ignoring the issues even when the evidence is right in front of them?  It’s as if they have repeated their own “performance is irrelevant” mantra so many times that they are actually starting to believe it.  It’s unfortunate, because Groovy really is an interesting effort.  I may not see any value for my needs, but I can understand how a lot of people would.  It fills a nice syntactic niche that other languages (such as Ruby) just miss.  But all of its benefits are for naught if it can’t deliver when it counts.

That did it.

(Editor's note: Shiiiiiiiit! I didn't say nothing, HE did, why're you swinging that beer stein at *WHACK*)

^^^^^^^^^^^^^^^^^^^^^

OK, now that we've gotten that out of our system, let's sit back and examine this issue more carefully, shall we?

The fact is, Groovy is slower than it should be. The Groovy guys can mumble about how performance isn't that important and that developer productivity is what really matters and similar kinds of rationale, but at the end of the day, the basic fact remains that Groovy is, by measurement of several different tests, at least an order of magnitude slower than compiled Java code.

Or is it? Funny thing is, looking at some of these tests, they don't say whether the Groovy code was compiled first, or run through the Groovy interpreter. Theoretically this shouldn't matter, since the Groovy architecture essentially compiles the classes generated once read, it might make a difference in practice.

Although Daniel's post doesn't mention it, I went back and double-checked. Peter Knego's benchmark says that "Groovy code was inside a Groovy script, compiled with groovyc.", which leads me to believe that it was compiled code rather than run through the Groovy shell interpreter, but it would be nice if the actual code and batch/command scripts that ran the benchmarks would be available. (He also notes that each time, the benchmark was "warmed up" by running the bechmark five or six times in a loop, presumably to allow the JIT to work its magic, but most notably, doesn't point out which JVM was used, -client or -server.) Meanwhile, Derek Young's ray-tracing example explicitly uses the Groovy interpreter, but defends that decision in comments: "The only reason I didn’t use groovyc was because the difference was so great, and the compilation overhead at the beginning of the run only takes a couple seconds. I decided it wasn’t worth waiting another two and a half hours to time the compiled output. Running with groovy first compiles the code just like groovyc does, then executes that code. It doesn't interpret the source code or run any differently."

So, apparently, it doesn't make much difference. That's not a good development for the Groovy language.

But let's put the hard numbers out of the way for a moment, and concentrate on the much bigger question: does the core Groovy team just not value performance? And, as a corollary to this, does the performance of a language really matter in a day and age when CPUs are still doubling in size and number of cores? Rick's (and others', including myself) positions on this seem fairly clear, that we long ago passed the threshold where programmer time became more expensive than CPU time, and therefore we should optimize based on programmer productivity, not CPU efficiency, and that's important to recognize: a language should enable the programmer to express the core idea without a great deal of "noise" or additional work, what Stu Halloway has coined as "ceremony", and certainly Groovy takes the Java programmer a step closer towards that place of lower ceremony.

But...

But I can't help it, folks. Ted's First Law of Computer Science states that "Dogma is the Root of All Evil", and holding scripting languages up as the last language you'll ever have to use is dogma, plain and simple. Ted's Second Law of Computer Science states, "Context matters", and in this case, the context includes the performance cost of using a language or tool. Taking a performance hit that weighs in at the orders of magnitude mark is just too big to ignore--the ray-tracing example, at its close-to-four-orders-of-magnitude hit, almost suggests that it would have been just as expensive to offload all those calculations through a distributed RPC call to another machine, rather than calculate it locally in Groovy, and when it becomes faster to go off-CPU to do a calculation than to do it locally, something is wrong.

And Daniel's point is good to hear clearly through the noise: "When you start getting hundreds of thousands of hits an hour, you’ll be willing to sell your soul for every last millisecond." Forget getting hundreds or thousands of hits an hour--the real test will be when the system gets hundreds or thousands of hits per second, that's when developers will be scrambling to find ways to eke out those last bits of performance from the system, even if it means selling their last Mountain Dew (which for some is pretty much synonymous to "soul") to whatever entity can give it to them.

So what exactly is my point with this particular entry, besides stirring the pot up a little? In order:

  1. Measure for yourself. As with all things performance and scalability related, abstract benchmarks aren't a good measure of how well it works for you and your system. Build a prototype, measure, and then compare that against your performance and scalability goals. You did establish performance and scalability goals as part of your project's runup, right? (If you didn't, then you probably assume your users don't care about performance, and I suspect you'll be rudely surprised on the veracity of that statement before long...)
  2. Benchmarks are tricky things. Programmers could learn something from politicians, and that is the imprecise nature of poll results. Thanks to the nature of statistical analysis and the sample size and source used to produce the poll, polls are always cited with a "plus or minus 3 percent" (or 5 percent, or 10 percent) to indicate the imprecision assumed in the poll. Benchmarks, both across languages and across other products, should be assumed to have similar kinds of imprecision. As people have already noted, benchmarks very quickly get "gamed" in order to produce results that are unfairly biased one way or another if they're not explicitly written and administered to be fair and equal to all sides involved. This isn't to say that benchmarks aren't useful, they're just not useful to a point more precise than rounding to the nearest 10% figure.
  3. Groovy IS slower than it could or should be. As much as I like the people involved in the Groovy space, and as much as I like the language itself, I can't help but be very very worried that Groovy's performance numbers aren't anywhere close to where they should be. Yes, productivity will get you a long way in the technology-adoption market, but once people have adopted your language or tool, if your system proves to be unresponsive and performance-challenged, the doors letting people in will get blocked by the people trying to get out, and that's not good for Groovy.
  4. Groovy's performance may not be a reason not to use Groovy. Let's be honest, again: productivity matters. This is Mort's principal goal, remember, and there's nothing wrong with it. Groovy fits into that most natural of places, a scripting language gluing together pieces written in a system language. Perl and Python serve the same purpose for native/C/C++ code, PowerShell does the same (I believe) for .NET code, and it's high time we see the value in doing that for Java code.
  5. I believe that the core Groovy team holds performance as a value. I know Graeme and Guillaume, and I believe they believe in the value of performance. I believe that Groovy will get faster over time, as they discover new and better ways to compile Groovy code into bytecode. That doesn't mean users of Groovy should walk into this exercise with their eyes shut, mind you, but take the whole of this discussion into context as you figure out where Groovy can be used to make your life, as a developer, more productive and powerful. Certain parts of your system are perf-sensitive, and certain parts aren't. Identify which of those parts are which, and apply Groovy (and other tools) judiciously.
  6. These discussions are always good, so long as they're held without rancor. Groovy doesn't suck. It has warts, but so does everything. Hushing them up or pooh-poohing them just leads to arguments--I encourage the Groovy team to take the criticism of Groovy's performance the way I intend it in this blog entry: a challenge to be faced and overcome, and not as an indictment of any and all Groovy code everywhere. Because, and I will say this outright, if Groovy's backers seriously mean Groovy as "a better Java 7", then they have a large gap to fill.

 

Oh, and if some of you wouldn't mind sticking around to clean up the mess...? Getting beer off the ceiling can be tricky.