Category: Performance

Fear of Streams

There’s an article in the new issue of the French magazine Programmez which got my attention. It chronicles the efforts of some programmers for a B2C site to investigate converting their old for loops into more readable and maintainable Java-8 streams. The article is kind of a mess, actually, but to a software craftsman or craftswoman who works on Java legacy code, the message is a bit frightening.  The programmers did performance tests of the methods they changed, but because some of the more complex methods they modified took twice as long to execute as they did before, their employer was not convinced and does not want them to proceed with a refactor.

If you thought that it was difficult to sell your product owner on the idea of adding days to the schedule to clean up legacy code, imagine having to reveal that the clean-up also involves the risk of performance issues!

Not all of their changes slowed down execution – some even speeded it up a bit.  What troubled me about the article, though, is that I could not predict, looking at the examples they gave of their modifications, which ones would run faster and which ones would run slower than before.

I turned to StackOverflow to see if somebody knows how to do that in general. Is there any way, in a code review, for example, to see if a stream will run significantly slower (a constant-order slowdown, but enough to annoy a user or kill a refactoring initiative) based on just looking at the code?  Of course, it’s possible to code stupidly in a way that will obviously cause a slowdown.  But, apart from some useful guidance about when to use parallel streams and when not to, the message I got from some smart, informed developers is that you can’t, or else it’s not worth the trouble for what amounts to at worst a constant-order slowdown. Of course you should review code changes to make sure they don’t affect the correctness or the complexity of the algorithm, and maybe you can tell whether going parallel would help or hurt performance, but that’s all you can usefully detect.

I think the biggest mistake that the developers in the Programmez article made was to evaluate the performance based on isolated unit tests (microtests).  Performance of a method is not interesting to the end-user.  A  constant-order decrease in performance of a method will frighten a product owner, but maybe the end-user won’t even notice it.  For example, it may be that doubling the execution time of a loop results in a 0.001% decrease in performance of a web request because it’s a loop to process results of a heavy database query.

If the code you work on is for a financial trading application where every nanosecond counts, then you’re not going to be using Java streams at all – you likely have customized collections that run super-fast with no syntactic frills.  For the rest of us, we need to do automated end-to-end performance tests for the scenarios where performance might be an issue.

If you convert for loops to streams, don’t convert everything on the same day. Space out these kinds of changes enough that end-to-end tests will catch only one performance issue at a time.  Then you can reassure the product owner that there will be no “noticeable” decrease in performance from the refactor, but that the code will be more readable and understandable, and thus easier to change and less bug-prone.  For certain iterations which can be parallelized efficiently, there may even be noticeable performance gains.

The String Thing: Java String Concatenation Performance

I’m researching an upcoming performance-related article, and it occurred to me that I had always just accepted it as given that concatenating Strings using the plus operator in a loop is bad.  First I used StringBuffer instead (back in the day), and then StringBuilder because StringBuffer is synchronized. Occasionally, I started using Guava’s Joiner (which has nice, clean syntax, but involves a 3rd-party dependency). Now Java 8 offers a few new approaches to String concatenation, the most eloquent of which is String.join().

So, rather than rest on received wisdom, I thought I’d put these different approaches to the test using JMH benchmarks.  The source code for my benchmarks is here. It concatenates a list of 20000 Strings into one underscore-delimited String in 7 different ways.

Here are the results (on an old windows 7 laptop with 4MB of overly-solicited RAM and 2 cores, so your mileage may vary) – lower scores are faster:

Benchmark                                                 Mode  Cnt       Score       Error  Units
concatenationUsingGuavaJoiner                             avgt   20      70.193 ±     2.456  ns/op
concatenationUsingPlus                                    avgt   20  231599.954 ± 20386.920  ns/op
concatenationUsingStringBuffer                            avgt   20      58.894 ±     3.943  ns/op
concatenationUsingStringBuilder                           avgt   20      52.580 ±     1.514  ns/op
concatenationUsingStringJoin                              avgt   20      69.401 ±     5.827  ns/op
concatenationUsingStringJoinerInStreamWithForEachOrdered  avgt   20      64.491 ±     3.789  ns/op
concatenationUsingStringJoinerOnForLoop                   avgt   20      69.301 ±     6.815  ns/op

Briefly, this shows that concatenation using ‘+’ is truly awful (more than 4300 times slower than the StringBuilder approach!). All other approaches are roughly equivalent in performance, though StringBuilder runs about 30% faster than the syntactically superior approaches. The biggest surprise for me was how little difference there is between StringBuffer and StringBuilder (StringBuilder is about 12% faster). I was also a bit surprised that the compiler/JVM doesn’t do a better job of optimizing the ‘+’-based concatenation, given that this is a known issue since the 1990’s (it turns out the received wisdom of old still applies).

So, if you must have speed at all costs, the winner is:

	public String concatenationUsingStringBuilder() {
		// The approved pre-Java-8 approach
		StringBuilder result = new StringBuilder();
		for (String s : sourceList) {
			if (result.length() > 0) {
		return result.toString();

Otherwise (if having readable, maintainable code is worth more in the long run than a minor performance gain in a few parts of the code), here’s the real winner going forward:

	public String concatenationUsingStringJoin() {
		// The simplest, most concise Java-8 approach
		return String.join("" + SEPARATOR, sourceList);

Addendum: Before someone adds a comment about it, let me say that for readability with String join, the code would be prettier if SEPARATOR were a String constant instead of a char constant.  Then you’d just have:

return String.join(SEPARATOR, sourceList);

I didn’t fix it in my benchmark because it’s a reminder that String.join() does not accept a character constant. Fixing it would have no effect on performance, which I can prove with  the output from javap -c on my benchmark class. It shows that the compiler has converted “” + SEPARATOR to the String constant “_”, which is loaded with the ldc bytecode).