6

We're trying to tweak some Oracle JVM garbage collection options and one developer tried to use -XX:PretenureSizeThreshold to make sure a large array of objects was put in Tenured right away. I'm pretty sure the assumption was that the array size equals or exceeds the total size of all the objects in it.

But in Java, aren't arrays of objects just arrays of references? I.e. each object in the array, as well as the array object itself, is separate in memory and treated as separate by the garbage collector? I think the array object can still get fairly large if there are millions of entries, but it shouldn't be anywhere near the total size of the objects it "contains" if each object is much bigger than a reference.

I think there's confusion because AFAIK, in C:

  1. It's possible to have an array of structs that really does store the structs.
  2. It's also possible to have an array of pointers to structs.

I'm pretty sure Java always uses 1. for arrays of primitive types and always uses 2. for arrays of objects, while C can use either for any type...?

What if I use an ArrayList with frequent append()s (as we are in the case at hand)? Is only the array copied, and not the objects in the array? Also, when the array is copied, even if the old array was in Tenured the new one starts in Eden, right?

Improve this question
1
  • You are basically correct. In Java an array of non-primitives is always an array of references. And if an ArrayList gets reallocated the new array starts "fresh". (This is not to say that some JVMs don't special-case certain scenarios, but that would be pretty much one-off and targeted at a specific (ie, benchmark critical) situation.) Commented Oct 16, 2013 at 19:12

3 Answers 3

Reset to default
5

But in Java, aren't arrays of objects just arrays of references?

Just references. All objects are allocated on the heap, never in arrays or on the stack (at least officially, the optimizer may use stack allocation if possible, but this is transparent).

it shouldn't be anywhere near the total size of the objects it "contains" if each object is much bigger than a reference.

Yes, in Java whenever you say "assign/store an object", you mean the reference (pointer in C terminology).

What if I use an ArrayList with frequent append()s (as we are in the case at hand)? Is only the array copied, and not the objects in the array?

The array gets only copied when resizing is needed, i.e., very rarely and the amortized cost is proportional to the number of inserts. The referenced objects gets never copied.

Also, when the array is copied, even if the old array was in Tenured the new one starts in Eden, right?

Yes!

Improve this answer
Sign up to request clarification or add additional context in comments.

4 Comments

Did you mean to quote the same thing twice? Still, you answered all the questions, so I accepted :)
Wow. This sounds awful, but perhaps Java's allocation is good enough to keep data adjacentcy at a maximum. If not.. well, that's a tragedy of lost performance.
@Thumbz: Sure, the allocator places object allocated in row next to each other (unless it itself has to get a new memory chunk). But then the compacting GC comes... it also tries to preserve locality, but it gets more complicated. Sometimes it may be a problem, but don't forget the root of all evil thingy.
Normally I'd agree with the premature optimization concept, but when it comes to day structures and the adjacentcy of data.. well, reading from a disk is slow. It's the bottleneck of most programming.. basically anything that isn't really heavy in synthesis... (I say with no authority - ISWNA)
2

Using -XX:PretenureSizeThreshold for tuning is unlikely to help you. This parameter applies only to direct Eden allocation, while most allocation is happening in TLAB (Thread Local Allocation Buffer) and -XX:PretenureSizeThreshold is ignored.

TLAB could be quite large for thread actively allocating memory (few megabytes).

You can tweak TLAB sizing, to reduce this effect, but that would probably do more harm than good.

Improve this answer

1 Comment

Wow. And I thought I'd dug into a good bit of GC's complexity already, only to google TLABs and run into this horrifying post. I wholeheartedly agree that at that level (and at most levels) leaving the VM to its own devices is the best bet. Thanks for the tip!
2

But in Java, aren't arrays of objects just arrays of references? I.e. each object in the array, as well as the array object itself, is separate in memory and treated as separate by the garbage collector?

Yes.

I think there's confusion because AFAIK, in C:

  1. It's possible to have an array of structs that really does store the structs.
  2. It's also possible to have an array of pointers to structs.

I'm pretty sure Java always uses 1. for arrays of primitive types and always uses 2. for arrays of objects, while C can use either for any type...?

Java, like C, typically stores arrays of primitive types as actual arrays with elements of those types. So an int[] array with 10 elements is typically going to reserve 10×4 bytes for the array, plus overhead for the entire array object.

Arrays of objects, however, are as you say, arrays of references. So an object[] of 10 elements is going to typically take up 10×4 bytes (or perhaps 10×8 bytes on 64-bit CPUs) for the array, plus overhead, plus space for each object that each non-null element references. This corresponds in C to an array of pointers.

(I use the term "typically", because even though that's how most JVMs do it, they are not required to allocate memory in any particular fashion.)

Also be aware that Java does not have true multi-dimensional arrays like C (or C#). An int[][] in Java is actually a one-dimensional array, where each element is a reference to its own int[] subarray. In C, an int[][] really is a two-dimensional array of integers (where the lengths of all but the first dimension must be known at compile time).

Addendum

Also note that, like you say, C can have true arrays of structs, which are neither primitive types nor pointers. Java does not have this capability.

Improve this answer

Comments

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.