MemTracer strikes back

April 5, 2008 – 9:59 am

You may remember my proof-of-concept experiment – MemTracer. It started as a hobby project, but recently I had chance to test it in a real world application. It quickly turned out there’s a big difference between my simple test program and full blown game performing hundreds of memory operations per second.

Below you can find list of conclusions I drew from this experience:

  • realtime analysis is next to useless. Things change too quickly anyway and it slows down the execution of client application significantly. You’ll want to do as little work in receiving thread as possible. Avoid switching threads/updating controls if you can. I just record memory operations for later analysis and update some really basic informations (on timer, not from receiving thread, Invoke() seems very slow). If you record every memory operation (alloc/free) and frame markers, you can recreate complete history easily at any given moment. All the more complicated structures are built on request, not while recording.
  • the simplest way to intercept all memory operations is to use Detours. I recommend starting with tracing only internal allocations, though (operator new/your own routines). Intercepting HeapAlloc and family causes some additional problems, because number of memory operations to handle skyrockets then, especially DirectSound seems to allocate like crazy. What’s more dangerous, sockets can allocate memory as well, so you’ll need some kind of mechanism to prevent neverending loops. Simple way to limit noise/signal ratio is to only consider memory blocks bigger than some threshold (say 1kb). Worked really well in my case, most of the low-level applications are parts of bigger structures anyway, so if you eliminate big leak, you’ll get the small ones as well. If really needed, after finishing with big leaks, we can track small blocks only. Divide & conquer, baby.
  • StackWalk64 is reliable, but slow. It may not be noticeable normally, but in places with big amount of memory applications (loading big levels for example) it really does slow program down. Use EBP/ESP walk or RtlCaptureStackBackTrace function when possible (you may need to fall back to StackWalk64 for some modules with FPO).
  • you’ll want to use some tricks to cut down your client application memory usage. After my first tries it tended to run out of memory on longer tests. Basic modifications I made to reduce memory footprint were:
    • tags were 4CC codes by default. If some block needed longer tag, I would only store CRC per block (same field as 4CC), real entry would go to global dictionary.
    • similar situation with callstack. Instead of keeping 20+ entries for every block, I only stored callstack CRC. This also speeds callstack comparisons up.
  • it is essential that your tool is 100% reliable. Nothing’s worse than false alarms, they can make you bark at the wrong tree for hours, only to find out everything was right, it’s bug in the tool. Make sure it works in every situation before adding fancy features. It took me some time, but now I’m at stage when if MemTracer reports something leaks — it leaks, no matter how improbable the callstack may seem. Simplifies life, really, because I dont have to double check every time. Make sure allocations are caught properly in the game, that every packet sent from the game reaches the tracer, that every allocation in the tracer is recorded properly.
  • last, but not least – make sure your sockets work in non-blocking mode (or run sending in another thread) :). Took me some minutes to figure out why the game suddenly slowed down to a crawl after I connected to it. It is a little bit more work (for example socket may not be always ready for sending more data etc), but it definitelly pays off. You may start with blocking sockets for quick’n’dirty start, of course.

I also extended MemTracer’s functionality. Many of the ideas employed were stole^H^H^H^H, err, inspired by aforementioned van der Beek’s article. I’ll show most interesting features here, so hopefully you can be “inspired” as well :>. Figure 1 shows application main panel. The only weird control here is filter combobox, used to track memory stats for various allocators, it lets to display statistics only for blocks with selected tag.

mainview
Figure 1: The main panel. List of snapshots + global stats.

Double-clicking any snapshot shows dialog with gory details. Before analyzing individual snapshots, however, we can display some general memory info. First, let’s see what block sizes are most commonly allocated. Entries can be of course sorted by any criterium. As you can see in figure 2, there’s also second tab, where biggest allocated blocks are shown.

aoverview
Figure 2: Allocation overview.

Another panel shows all places in application with number of blocks/bytes allocated by them:

allocpoints
Figure 3: Allocation points.

Double clicking any entry jumps to corresponding place in the source code (file/line). These 2 windows give basic information about your memory usage patterns and can already help in finding unwanted situations like per-frame function allocated too many blocks every call. First dialog can be especially useful in finding what block sizes should use a free list allocator.

snapshot
Figure 4: Snapshot details

Double-clicking snapshot entry, either in the list or in the right summary pane opens another window with more detailed information. It contains complete mem block tree for selected snapshot (double-click jumps to proper location in the source code, as in every other dialog). Tree can be sorted top-down (default) or bottom-up (very useful in finding critical paths in the program, allocating most of the memory). Additionally, blocks can be separated basing on their tags. Consider figure 5, where we show tagged overview of heap allocations. Every low-level heap allocation is tagged with high 16 bits of its heap address. Entries in every mode are sorted in respect to number of allocated bytes, so it’s trivial to see what heap is allocating the most memory.

tagged
Figure 5: Tagged view of snapshot

At every branch of the mem block tree it’s possible to show view of the subtree originating at selected branch. It’s the most detailed view showing individual blocks. This mode is especially useful in connection with tags.

Tag is either 4CC or string up to 32 characters. Block can be either tagged at allocation time or any moment later, via separated message. It’s very useful for identifying resources for example. Tagging blocks is easy when tracking only internal allocations, becomes a little harder when intercepting all calls, because for example single CreateTexture() call may result in multiple HeapAllocs, we’d like to tag all the blocks with texture name obviously. In order to achive this, I introduced a concept of “tag stack”. If it’s not empty, every allocated block is tagged with entry from top of the stack. So, you push texture name before creating it, pop after and it tags all the blocks allocated between those calls.

subtree
Figure 6: Subtree details for all blocks originating from CExoSoundSourceInternal::Demand().
Tags are resource names.

Example is shown in figure 6. I was investigating a potential leak, where at the same point in game, after first run there were 3 blocks allocated by shown path, and the next time – 4. As we can see, it’s a false alarm, we just were “lucky” enough to record the snapshot in the exact moment we received a new quest (that’s our 4th allocation, not present in the original snapshot). Screenshot is a little outdated, later I added possibility to filter subtree entries basing on their ID#. It makes easy to display only allocations “older” than specified ID. Scroll back to figure#1, as you can see after memory operation #761718 level should be unloaded. We can filter out allocations “newer” than that and see if there are any leaks left, even see addresses/tags/size of individual leaking blocks.

snapshotcompare
Figure 7: Snapshot comparison.

There are two modes of snapshot comparison: difference and overlap. Difference is “standard” mode, when third tree consists of entries that are present in snapshot B, but arent in snapshot A. It’s usually applied to comparing snapshots taken in the same “location” in game, just in different moments. Canonical example is main menu: take snapshot, load level, play a little bit, go back to menu, take snapshot. In the ideal situation both snapshots should be identical. Second comparison mode is called “overlap” and it shows exactly that – blocks that are present in both snapshots. Typical application would be to load a level, take snapshot, play a little bit, move to another level, take another snapshot. Display common blocks, see if some shouldnt be there (for example, enemy only present at the previous level). It’s powerful tool for finding logical memory leaks. Important distinction between comparison modes is that “difference” compares callstacks/sizes, while the actual addresses dont have to be equal, while “overlap” only considers blocks equal if they’re exactly same blocks.

MemTracer contains also bonus feature – traced variables. Application can send current value of every variable it wants to trace, every time it changes. Typical candidates would be variables like number of objects, number of models, textures and so on. It makes spotting connections easier. Number of game objects keep growing and growing during gameplay? Something’s wrong, let’s take closer look. Sound cache size doesnt seem to be limited? May need further investigation.

tracedv
Figure 8: Traced variables

Current version of MemTracer is in certain areas less advanced than my first try, as I said, I focused on stripping stuff slowing application down. There’s no frame-by-frame analysis yet and no fancy usage graph. I dont really miss the graph, but I plan to add frame analysis in the future as it may be helpful in finding some really obscure bugs. However, even without such fine-grained analysis, the application already proved useful and helped in finding some memory-related problems that I probably wouldnt find otherwise. Whole system really isnt hard to code, the client part was done in a week or so, after hours, at home (and I really suck at C#, it’s my first app ever). Much time went in the server part, especially fighting with Detours in multi-threaded environment, but chances are – you wont need it at the beginning.

I dont really plan to extend MemTracer any further. It started as an experiment, morphed to something semi-usable and served its purpose. Modifications that would need to be made in order to make it more universal:

  • support for DLL injecting. Right now the application has to include custom C++ code and call some functions. DLL injecting would allow to examine third party executables. Didnt bother to implement it as I wanted only to check my code anyway. Plus, the advantages are biggest if you can provide some extra information about memory usage (like tags).

  • heap view I mentioned in the previous article, still not done. Problem is, while it could be rather easily done for user managed sub-heaps, to visualize it for system heap would most probably require crazy amounts of memory (need to store heap state for every frame… perhaps could be compressed by storing deltas only, something to consider).

  • ability to “import” and combine information from multiple sessions. It’d show snapshots from all imported sessions and allow for normal operations on them (compare, details etc).

 

  1. 3 Trackback(s)

  2. Sep 2, 2008: realtimecollisiondetection.net - the blog » Posts and links you should have read
  3. Jan 18, 2009: Lightning Engine » Blog Archive » Tracking Memory Allocations
  4. Nov 14, 2010: MemTracer – episode IV | .mischief.mayhem.soap.

Post a Comment