Slashdot Mirror
Help crack the Java 1.6 Classfile Verifier
pdoubleya writes "As part of the development of Mustang (Java 1.6), Sun is developing a new, smaller and faster classfile verifier which they want your help in trying to break. As Sun VP Graham Hamilton puts it in his blog entry, "As part of Mustang we will be delivering a whole new classfile verifier implementation based on an entirely new verification approach. The classfile verifier is the very heart of the whole Java sandbox model, so replacing both the implementation and the basic verification model is a Really Big Deal.... The new verifier is faster and smaller than the classic verifier, but at the same time it doesn't have the ten years of reassuring shakedown history that we have with the classic verifier." You can read about the new verifier on Gilad Bracha's blog, and join the new Crack the Verifier initiative to if you can break it. Read all about the Crack the Verifier - Challenge."
If his name doesn't ring a bell, he's a Java guru who works for Sun and wrote the 2nd and 3rd editions of the Java Language Spec. A bunch of his papers are listed here.
It's a relief that JDK 1.6 won't include any language changes (as far as I know?). Updating various parsers and whatnot to work with all the JDK 1.5 language changes was a big job, although some of the new features certainly are quite handy.
The Army reading list
This will allow the vm to do small amounts of gc whenever it has a chance, as opposed to wating until an allocation request will fail and then running through the entire heap.
Our app runs 24/7 and while the markets are open and 10meg+ of raw data is coming down the line every second, we can't allow the app to take a timeout for a gc run. The app runs in 256meg, too.
Java will ALWAYS keep the 64MB heap. It will NEVER shrink it.
Who cares? This is completely irrelevant, as long as the JVM marks the pages it's not using as discardable, which modern JVMs on modern OSes do.
You have to remember that all memory is virtual. I can allocate a 1GB array, but as long as I never actually touch the pages (making them "dirty"), no storage, whether RAM or disk, is ever used. When the JVM allocates its 64MB, that virtual memory is initially all "clean" and therefore consumes no RAM. As it's used, it gets dirtied and physical RAM gets mapped to it... but when a garbage allocation occurs, both the Sun and IBM JVMs mark the now-unused pages as clean, allowing the OS to reuse them at will. Effectively, they no longer consume any space.
Even if the JVMs didn't mark the pages as clean, the impact of the JVM holding onto the 64MB wouldn't be that significant. The allocated, dirtied but now-unused memory will simply get swapped out to disk, allowing those pages of RAM to be used by other applications.
With a decent OS, it really doesn't matter if an app holds onto some memory that it's not using, especially if it has the decency to tell the OS that it's not actually using it.
That said, there is a "problem" here, it's just not the one you're pointing out. The problem, if you want to call it that, arises from the generational, copying, compacting garbage collector used by modern VMs. Now, don't get me wrong, this GC is very cool. It's significantly more efficient than typical malloc/free memory management, *and* it eliminates some major classes of memory leaks (programmers can still leak memory with GC, but it's harder).
Without getting into the details, although GC is safer and more efficient in terms of CPU cycles, and although on average it doesn't use a great deal more memory than manual allocation would use (particularly since many performance-sensitive manual allocation apps end up managing their own memory pools in order to avoid the run-time cost of malloc() and free()), GC does tend to increase the number of memory pages that get touched over time.
Why? Two reasons. First, suppose the application allocates a small chunk of memory, uses it, frees it, allocates another small chunk (about the same size), and then uses and frees it. Most malloc()/free() implementations will tend to return the same chunk of memory for both allocations. Repeating the process a thousand times (which isn't all that uncommon) will probably only dirty a single page of memory. With the sort of garbage collector used by current JVMs, every one of those allocations will return a different chunk of memory, and many pages will therefore be dirtied. In terms of CPU cycles, GC wins big, because, rather than a thousand small frees, there is one big one. And allocation is up to two orders of magnitude faster. It may sound like the GC approach is going to use a lot more memory, on average, but it's not that bad because of the tendency of malloc/free-managed heaps to become fragmented. On average, GC implementations don't have many more pages in use than malloc/free implementations, and may actually have less, but the GC allocators tend to "roam" across the pages.
Second, the "copying" nature of the GC. The main thing that makes generational, copying GC-based memory management so fast it that it never has to deal with fragmentation. To describe it in an oversimplified way: Every now and then, the GC relocates all of the in-use objects into a nice, compact block. That makes allocation fast, and tends to reduce the number of pages in active use, but it increases the number of pages that get dirtied and therefore require actual RAM to be provided by the OS. The copying also has a cost in CPU cycles, but it's small relative to the cost of managing and searching free lists, which is what malloc/free implementations have to do.
Theoretically, as the GC copies objects and marks the pages where they used to live as discardable, the OS coul
Note to ACs: I usually delete AC replies without reading them. If you want to talk to me, log in.