During my first months as a PhD student (back in Feb 2005) at AUEB, we had frequent technical arguments with my then supervisor, Diomidis Spinellis, regarding the execution speed of various programming languages. Diomidis's argument was that the design of the Java language and the JVM were inherently less efficient than that of natively compiled languages, as they force upon us services that we may not want to use. The prominent example was garbage collection. To prove his point, he setup a simple experiment, which he documented in this blog post. In the experiment, he created random integers which he stored in an always-sorted container (TreeSet in Java, STL set in C++).

Back then, the result came to me as a surprise. Despite my best efforts (I did try a lot of VM options), I could not manage to bring the Java implementation's performance anywhere near C++. During the 7 years in the between, I occasionally run the code on any new system that I had at my hands, and the result was, give or take, the same.

Today, I decided to rerun the experiment. I compiled the C++ code with Clang 3 with all optimisations enabled (-O3 -march=corei7) and used the 1.6.0_31 version of the JVM to run the Java code. The result came to me as a surprise:

$ clang++ -O3 -march=corei7 sort.cpp
$ time ./a.out
real    0m1.063s
user    0m1.026s
sys 0m0.035s
$ javac SortInt.java
$ time java SortInt
real    0m1.102s
user    0m2.325s
sys 0m0.137s
$ time java -server SortInt
real    0m0.866s
user    0m1.068s
sys 0m0.071s

The Java version was faster than C++! Why did this happen? It turns out that between those 7 years JVM performance engineers did not sit idle:

  • Escape analysis is turned on by default after version 17 of the server Hotspot compiler (I was running version 20.6). It allows the compiler to analyse whether objects escape the context of a method and, if not, to remove locks and allocate them on the stack. Consequently, this reduces the load to the garbage collector. The difference escape analysis presumambly makes can be seen by running the JVM in verbose GC mode using the -verbose:gc -XX:+PrintGCDetails options. In the case of the server VM, only one minor collection is required!

  • There were significant improvements on the garbage collector section as well. The young generation collector is now parallel by default; this is one of the advantages of having multiple CPUs on modern machines. The full heap collector is also parallel by default in all collection phases, leading to shorted pauses.

  • The compiler performs supposedely significantly better register allocation on machines with many registers.

The Java platform is a prime example of how research is being put into practice. All features described above (and others: lock coarsening, biased locking, thread local heaps) were published papers in major software conferences (OOPSLA, PLDI etc). What is perhaps more interesting is that it is not just Java that benefits from the work being done on the JVM; Scala and Ruby, the two other languages mostly associated with the JVM benefit too. For example, in Java 1.7, the new invokedynamic opcode allows JRuby to optimize execution various dynamic execution aspects delivers significant performance improvements. In all, Java as a platform does seem like a healty development target; I am so sure about Java as a language.

To sum up: is Java inherently slower? Yes, it is, but hard optimization work has lifted several performance hurdles. Does it matter? To some problem domains, it does; I would never think writing the code that processes big data in Java, expect if distribution could lead to significant speedups. Most of the problems I am trying to solve are better expressed in Ruby and Scala. I am happy as long as those two languages offer 80% of the performance of Java.


14 May 2012



Submit to HN