Garbage Collection: Theory and Practice

Android’s VM relies upon an automatic memory management scheme; this cornerstone of the programming paradigm on which Java is based on has been a part of the Android ecosystem since its inception. For anybody unfamiliar with the concept, a quick explanation is that a programmer is neither in charge of having to either manually to reserve (allocate) physical memory nor freeing this memory up by himself once he is done with it. This is an important distinction to lower level programming languages where manual memory management is the norm. Of course, the advantage of automatic management is that it saves the developer from having to worry about memory management. The disadvantage is that the developer also no longer has control, and is at the whim of the system to correctly manage things in an optimal manner.

Android and Dalvik have traditionally suffered a lot from Dalvik’s garbage collection (GC) mechanism. Each time an application needed memory to be allocated and the heap (a space of memory dedicated to that app) would not be able to accommodate that allocation, the GC would fire up.

A GC’s job is to traverse the heap, enumerate all objects allocated by the application, mark all reachable objects (meaning, mark all objects that are still in use), and to free up whatever is left remaining.

In Dalvik’s case, this process resulted in two pauses; one during the enumeration phase, and one during the marking phase. A pause in this sense is that all code execution is halted for the totality of the application on all its threads. If the pauses are too big, this would result in dropped frames in the rendering of the application, which in turn resulted in jank and stuttering in the user experience.

Google claims the average duration of these pauses went up to 54ms on a Nexus 5 device, which results in a minimum of 4 dropped frames on average each time the GC would activate.

In my own experience and investigating, this number can blow up dramatically depending upon the application. For example, the official FIFA app is a poster child case in which garbage collection can run wild;

07-01 15:56:14.275: D/dalvikvm(30615): GC_FOR_ALLOC freed 4442K, 25% free 20183K/26856K, paused 24ms, total 24ms

07-01 15:56:16.785: I/dalvikvm-heap(30615): Grow heap (frag case) to 38.179MB for 8294416-byte allocation

07-01 15:56:17.225: I/dalvikvm-heap(30615): Grow heap (frag case) to 48.279MB for 7361296-byte allocation

07-01 15:56:17.625: I/Choreographer(30615): Skipped 35 frames! The application may be doing too much work on its main thread.

07-01 15:56:19.035: D/dalvikvm(30615): GC_CONCURRENT freed 35838K, 43% free 51351K/89052K, paused 3ms+5ms, total 106ms

07-01 15:56:19.035: D/dalvikvm(30615): WAIT_FOR_CONCURRENT_GC blocked 96ms

07-01 15:56:19.815: D/dalvikvm(30615): GC_CONCURRENT freed 7078K, 42% free 52464K/89052K, paused 14ms+4ms, total 96ms

07-01 15:56:19.815: D/dalvikvm(30615): WAIT_FOR_CONCURRENT_GC blocked 74ms

07-01 15:56:20.035: I/Choreographer(30615): Skipped 141 frames! The application may be doing too much work on its main thread.

07-01 15:56:20.275: D/dalvikvm(30615): GC_FOR_ALLOC freed 4774K, 45% free 49801K/89052K, paused 168ms, total 168ms

07-01 15:56:20.295: I/dalvikvm-heap(30615): Grow heap (frag case) to 56.900MB for 4665616-byte allocation

07-01 15:56:21.315: D/dalvikvm(30615): GC_FOR_ALLOC freed 1359K, 42% free 55045K/93612K, paused 95ms, total 95ms

07-01 15:56:21.965: D/dalvikvm(30615): GC_CONCURRENT freed 6376K, 40% free 56861K/93612K, paused 16ms+8ms, total 126ms

07-01 15:56:21.965: D/dalvikvm(30615): WAIT_FOR_CONCURRENT_GC blocked 111ms

07-01 15:56:21.965: D/dalvikvm(30615): WAIT_FOR_CONCURRENT_GC blocked 97ms

07-01 15:56:22.085: I/Choreographer(30615): Skipped 38 frames! The application may be doing too much work on its main thread.

07-01 15:56:22.195: D/dalvikvm(30615): GC_FOR_ALLOC freed 1539K, 40% free 56833K/93612K, paused 87ms, total 87ms

07-01 15:56:22.195: I/dalvikvm-heap(30615): Grow heap (frag case) to 60.588MB for 1331732-byte allocation

07-01 15:56:22.475: D/dalvikvm(30615): GC_FOR_ALLOC freed 308K, 39% free 59497K/96216K, paused 84ms, total 84ms

07-01 15:56:22.815: D/dalvikvm(30615): GC_FOR_ALLOC freed 287K, 38% free 60878K/97516K, paused 95ms, total 95ms

Above is a log extract of the first few seconds after launching the application. The garbage collector was invoked for a total of 9 times, causing the application to freeze for a total of 603ms, and a total of 214 dropped frames. Most of the pauses are caused by requests for memory allocation, as described by the “GC_FOR_ALLOC” tag in the log entries.

What ART promises and provides is a huge revamp of the whole garbage collection system, to virtually eliminate cases as the above. Here’s a comparison of the same situation under ART:

07-01 16:00:44.531: I/art(198): Explicit concurrent mark sweep GC freed 700(30KB) AllocSpace objects, 0(0B) LOS objects, 792% free, 18MB/21MB, paused 186us total 12.763ms

07-01 16:00:44.545: I/art(198): Explicit concurrent mark sweep GC freed 7(240B) AllocSpace objects, 0(0B) LOS objects, 792% free, 18MB/21MB, paused 198us total 9.465ms

07-01 16:00:44.554: I/art(198): Explicit concurrent mark sweep GC freed 5(160B) AllocSpace objects, 0(0B) LOS objects, 792% free, 18MB/21MB, paused 224us total 9.045ms

07-01 16:00:44.690: I/art(801): Explicit concurrent mark sweep GC freed 65595(3MB) AllocSpace objects, 9(4MB) LOS objects, 810% free, 38MB/58MB, paused 1.195ms total 87.219ms

07-01 16:00:46.517: I/art(29197): Background partial concurrent mark sweep GC freed 74626(3MB) AllocSpace objects, 39(4MB) LOS objects, 1496% free, 25MB/32MB, paused 4.422ms total 1.371747s

07-01 16:00:48.534: I/Choreographer(29197): Skipped 30 frames! The application may be doing too much work on its main thread.

07-01 16:00:48.566: I/art(29197): Background sticky concurrent mark sweep GC freed 70319(3MB) AllocSpace objects, 59(5MB) LOS objects, 825% free, 49MB/56MB, paused 6.139ms total 52.868ms

07-01 16:00:49.282: I/Choreographer(29197): Skipped 33 frames! The application may be doing too much work on its main thread.

07-01 16:00:49.652: I/art(1287): Heap transition to ProcessStateJankImperceptible took 45.636146ms saved at least 723KB

07-01 16:00:49.660: I/art(1256): Heap transition to ProcessStateJankImperceptible took 52.650677ms saved at least 966KB

The difference between ART and Dalvik couldn’t be bigger, as the new runtime manages to only pause for a total of 12.364ms over 4 foreground GC calls and two background GC calls. The heap did not need to grow at all during the application launch, in comparison to Dalvik’s 4 increases. The number of frames dropped diminished to 63.

Obviously this is a worst-case scenario of a badly developed application, as the application even in ART’s case still managed to drop a little over a second of frames, but bad programming practices such as overloading the UI thread is something that Android has to deal with on a regular basis.

ART offloads some of the work that is usually done by the garbage collector to the application itself, this eliminates the need for the execution pause during the enumeration phase. The second pause is vastly reduced by trying to do the work before the pause - a technique called packard pre-cleaning is applied, and the pause itself is reserved to a simply checking and validation of the work done. Google promises that they have managed to reduce these pauses to an average of 3ms, a good order of magnitude lower than presently achieved by the Dalvik GC.

The introduction of a special “Large Object Space” which is separate of the main heap, but still residing in the application memory, is destined to facilitate better handling of big objects such as bitmaps. These large primitive objects would cause great problems in terms of fragmenting the heap, causing more need for a GC invocation in case a new object needed to be allocated, but no continuous memory region big enough for it would be available. The frequency of GC invocations is thus greatly reduced due to less fragmentation of the memory heap and more intelligent allocation and elimination by the runtime.

Another good example is the launching of the Hangouts application; In Dalvik we see several GC pauses due to allocation calls:

07-01 06:37:13.481: D/dalvikvm(7403): GC_EXPLICIT freed 2315K, 46% free 18483K/34016K, paused 3ms+4ms, total 40ms

07-01 06:37:13.901: D/dalvikvm(9871): GC_CONCURRENT freed 3779K, 22% free 21193K/26856K, paused 3ms+3ms, total 36ms

07-01 06:37:14.041: D/dalvikvm(9871): GC_FOR_ALLOC freed 368K, 21% free 21451K/26856K, paused 25ms, total 25ms

07-01 06:37:14.041: I/dalvikvm-heap(9871): Grow heap (frag case) to 24.907MB for 147472-byte allocation

07-01 06:37:14.071: D/dalvikvm(9871): GC_FOR_ALLOC freed 4K, 20% free 22167K/27596K, paused 25ms, total 25ms

07-01 06:37:14.111: D/dalvikvm(9871): GC_FOR_ALLOC freed 9K, 19% free 23892K/29372K, paused 27ms, total 28ms

We have a good extract of all GC cases in the above log. The explicit and concurrent GC calls are the general cleanup and maintenance calls of the GC. The for_alloc calls are cases where the memory allocator is trying to allocate something but it doesn’t fit in the heap, and a GC is launched to try to make space. In the middle we see that the heap was expanded due to fragmentation and not being able to hold a larger object. The total “dead time” ends up at 90ms. In comparison, here’s the current ART extract in the L preview:

07-01 06:35:19.718: I/art(10844): Heap transition to ProcessStateJankPerceptible took 17.989063ms saved at least -138KB

07-01 06:35:24.171: I/art(1256): Heap transition to ProcessStateJankImperceptible took 42.936250ms saved at least 258KB

07-01 06:35:24.806: I/art(801): Explicit concurrent mark sweep GC freed 85790(3MB) AllocSpace objects, 4(10MB) LOS objects, 850% free, 35MB/56MB, paused 961us total 83.110ms

We’re not quite sure on what the heap transition lines represent, but we’ll assume they’re the heap resizing mechanisms. The only GC call is after the app is launched already and comes at a minuscule 961µS. We don’t see any of the previous calls to the GC. What is interesting, is the LOS statistics. We see that we have 4 large objects at 10MB in the LOS, allocation that would have otherwise been resided in heap space now totally foregoes it, and with that, avoids the repeated for_alloc GC calls and heap fragmentation that plagued Dalvik.

The memory allocation system itself was also revamped. While ART itself provided an about 25% boost over Dalvik, Google was not happy with this and introduced a new memory allocator replacing the currently used “malloc” allocator.

This new allocator, “rosalloc” or Runs-of-Slots-Allocator, was developed in mind for the usage cases of multithreaded Java applications. The new allocator has a more fine-grained locking mechanisms which can lock on individual bins of objects instead of the whole memory allocation itself. Small objects in thread local regions are able to completely forego locking altogether.

The result is a huge improvement in allocation speed, resulting in up to a 10x gain.

The garbage collection algorithms themselves have been reworked to improve the user experience and to avoid interruptions of the application. These algorithms are still a work in progress and Google has only recently introduced a new dedicated algorithm, the “Moving garbage collector” whole sole purpose is to defragment the application heap while applications are in the background.

Introduction and Architecture 64-Bit Support and Conclusion
Comments Locked

136 Comments

View All Comments

  • johncuyle - Wednesday, July 2, 2014 - link

    It's not a great reason to discourage the NDK. Many applications are written to be cross platform and run successfully on multiple architectures. It doesn't even increase your test load, since even if you're writing your application in Java you still do need to fully test it on every platform. Test is usually the expensive part. The exact wording of the page is just odd. It says that preferring C++ isn't a good reason to write your application in C++. That's a pretty obviously false assertion. Preferring C++ is a great reason to write your application in C++.
  • Exophase - Monday, July 7, 2014 - link

    Not directly, since this applies to DEX which never had a compatibility issue. But it might convince some app developers to stop using NDK due to the improved performance of DEX binaries.
  • johncuyle - Tuesday, July 1, 2014 - link

    The frame drop counts seem very odd with respect to the total milliseconds delayed. (Or I'm bad at math.) At 60 FPS a frame is 16 ms. A 4ms GC sweep might drop a single frame at 60fps. The output indicates it dropped 30 frames. That's 750fps. Plausible if you're running without vsync or framerate limiting on a static screen like a splash screen, but that's not really a meaningful example, nor is it especially noticeable to the end user. More interesting would be the frequency of a frame drop in an application with extensive animation running at an average of 30fps. That's going to be a situation where you notice every frame drop.
  • kllrnohj - Tuesday, July 1, 2014 - link

    It's a mistake in the article. Those log lines have nothing to do with each other. The Choreographer is reporting a huge amount of dropped frames because <unknown> took a really long time on the UI thread, *NOT* because specifically the GC took that time. This is actually pretty normal, as when an application is launched the UI loading & layout all happens on the UI thread, which the Choreographer reports as "dropped" frames even though there wasn't actually any frames to draw in the first place as the app hadn't loaded yet. So the 30 dropped frames there means the application took about 500ms to load, which isn't fantastic but it's far from bad.
  • Stonebits - Tuesday, July 1, 2014 - link

    "Overhead such as exception checks in code are largely removed, and method and interface calls are vastly sped up"

    This doesn't make sense to me--are exceptions handled some other way, or do you just keep executing?
  • extide - Tuesday, July 1, 2014 - link

    If an exception is not handled, your application crashes.
  • Gigaplex - Thursday, July 3, 2014 - link

    If the compiler can statically prove that a given piece of code won't throw, there's no need to insert the exception handling support. Not all exceptions will be removed by the compiler though.
  • Impulses - Tuesday, July 1, 2014 - link

    Hah, that PBT link was pretty funny... PBT FTW
  • hackbod - Wednesday, July 2, 2014 - link

    Re: "Because ART compiles an ELF executable, the kernel is now able to handle page handling of code pages - this results in possibly much better memory management, and less memory usage too."

    There shouldn't really be any difference. Dalvik was carefully designed so that its odex format could be mmapped in to RAM, allowing the kernel to do the same page handling as with ELF executables. (Actually a little better than regular ELF executables, since odex doesn't need any relocations that cause dirty pages.)

    The "ProcessStateJankPerceptible" and "ProcessStateJankImperceptible" are the process coming in and out of the foreground. When it goes out of the foreground, the GC switches to a compacting mode that takes more time to run but saves more RAM. When it switches back to the foreground, it switches to the faster GC you have been looking at. The GC pauses here won't cause any jank, because these switches are never done when the app is on screen.
  • tuxRoller - Wednesday, July 2, 2014 - link

    KSM mostly kicks in on virt loads, but the ksmd overhead is so small that including it on memory starved, multiprocess devices isn't a bad idea.
    BTW, I believe that android use the dalvik cache partition to avoid unnecessary re-jitting, but the space is limited, and therefore dynamic, so apps can be vacated.

Log in

Don't have an account? Sign up now