Thoughts on software development and other stuff

NFJS Boston Day 2

with 2 comments

Here are my highlights of No Fluff Just Stuff Day 2:

Garbage Collector Friendly Programming

Garbage collection is an interesting topic to me for several reasons, the main reason being that poor GC performance is very harmful in distributed environments. When you have a peer to peer system such as Coherence, any one node can directly communicate with any other node at any point in time to service a request. If a node that needs to service several requests is in a long GC pause, it isn’t just that node that is affected. Every JVM that is waiting for a response from that node also experiences high latency, thus causing a cascading effect. (More on what to do about this later.)

Brian Goetz described the evolution of GC in Java, starting with the single threaded mark and sweep algorithms up to the modern generational collector. What they each have in common is the tracking of allocation roots (which include static variables and local variables allocated on thread stacks) and the traversal of object references starting at these roots. The implementation of the generational collector is (roughly) as follows:

  • New objects are allocated in the young generation space
  • When a minor GC occurs, objects that have references pointing to them are copied to the survival space. The remaining objects are removed
  • Eventually objects that live long enough in the survival space are moved to the old generation.
  • If a minor collection fails to free up enough space in the young generation area, a full collection (which is much more expensive) will occur in the old generation space.

An implementation detail is that the JVM must track references from the old generation to the young generation in order to know which GC roots to traverse when performing a minor collection. This means that the more “old” objects there are pointing to “new” objects, the more work the collector has to do. In a practical sense, this means that allocating new objects is preferred to reference field updates.

It took a while to wrap my head around that last statement, so let me attempt to demonstrate. Let’s say I have a map that contains objects with many fields. If I wanted to update some of those fields I could do it like this:

Map map = ...
MyObject o = map.get(key);

Or I could do it like this:

Map map = ...
MyObject oOld = map.get(key);
MyObject oNew = new MyObject(f1, o.getField2(), f3, o.getField4(), f5);
map.put(key, oNew);

The first example updates three reference fields (assuming the object in the map is in the old generation), whereas the second is updating one (the reference held by Map.Entry) – and as an added bonus the second implementation is thread safe (assuming that MyObject is immutable and the Map implementation is also thread safe.) If anyone has a better (or more correct) example of this concept, I’d be happy to see it!

Many other concepts were covered including: why you should use finally instead of finalizers to clean up, weak references, soft references, and tracking down memory leaks. Capturing heap dumps to track memory usage is a technique I recommend to customers (this works much better than speculating/wild guesses about where unexpected memory allocation is coming from) – I especially recommend configuring the JVM to generate a heap dump when an OutOfMemoryError is thrown. My favorite tool to read heap dumps is Eclipse MAT. Heap histograms are also a nice light weight approach to analyze memory problems.

When customers ask for suggestions on GC tuning, my recommendation is to keep it simple: fixed size 1 GB heaps (on the Sun VM), and don’t use more than 75% of the heap. I usually don’t recommend any specific tuning parameters, as the GC algorithms are constantly improving and any exotic flags that may have worked in older JVMs (assuming they helped in the first place) may not work so well in newer JVMs. The best advice I can give is to not fill up the heap as this will cause more frequent full collections.

Inside the Modern JVM

NFJS tends to cover languages in the JVM other than Java (such as Groovy, Scala, JRuby, etc.) This is a testament to the strength and viability of the modern JVM. Brian covered some of the advancements and (quite frankly) rocket science that goes into the JVM, HotSpot in particular. No matter what happens to Java (which isn’t going away any time soon), the JVM will be around for a very long time.

In a nutshell: why is Java, a supposedly interpreted language, faster than C++ in many cases? The answer is that the JVM determines which optimizations to make at run time instead of compile time, which is the opposite approach of C++ and other native languages. Optimization at runtime is far more effective, since the JVM has hard statistics of real world usage to draw on as opposed to the speculation and guessing that happens when everything has to be compiled to machine code before execution.

The overall theme of this talk (and the previous ones) is to write simple clean code – the runtime recognizes common usage patterns in Java and is built to optimize these patterns.

Java Collections

Ted Neward gave an engaging and entertaining talk on the Java Collections API. To be honest I was familiar with most of the material, but he is a fun speaker to watch, in spite of the fact that he gave me a good ribbing for showing up to his talk after it had already started! He is quite biased against arrays and towards collections, which made me think back to a web/remote services API I designed a few years ago. I exclusively used arrays as the collection type for this API for two reasons:

  1. To make SOAP/cross language interoperability much simpler (least common denominator – every language does arrays)
  2. There were no generics at the time; using arrays instead of generics in the interfaces meant that I could explicitly define the type of the array

The second item is not as important anymore now that we have generics, so I’m inclined to agree that arrays should be used sparingly nowadays.

Another interesting point is that iteration over collections should be done using a functor instead of a plain iterator. For example:

List<Name> names = ...; 
MyListOps.apply(new MyApplyFn<Name>() { 
  public void apply(Name n) { 
    // use n 
}, names);

This allows the possibility of processing the collection in multiple threads.

What’s coming in Java 7

This was Ted Neward’s next talk which was just as interesting and opinionated. Here are the highlights:

  • The release is targeted for early 2010
  • There is no official JSR for Java 7
  • Most of the information on what is going into Java 7 can be found on the blog of Joe Darcy of Sun.
  • Alex Miller has a huge page on his blog detailing what is in and what is out. This is information that is is aggregating off the web.

One of the most compelling additions to Java 7 is JSR 292, which introduces the bytecode invokedynamic. The implications of this addition are narrated by Charles Nutter of JRuby. There are other syntactic conveniences making it in; however it will not include closures (a fairly controversial topic.)

Written by Patrick Peralta

September 12th, 2009 at 10:25 pm

Posted in Development