Monday, July 14, 2014

Notes from Calligra Sprint. Part 2: Memory fragmentation in Krita fixed

During the second day of Calligra sprint in Deventer we split into two small groups. Friedrich, Thorsten, Jigar and Jaroslaw were discussing global Calligra issues, while Boud and me concentrated on the performance of Krita and its memory consumption.

We tried to find out why Krita is not fast enough for painting with big brushes on huge images. For our tests we created a two-layer image 8k by 8k pixels (which is 3x256 MiB (2 layers + projection)) and started to paint with 1k by 1k pixels brush. Just to compare, SAI Painting Tool simply forbids creating images more than 5k by 5k pixels and brushes more than 500 pixels wide. And during these tests we found out a really interesting thing...

I guess everyone has at least once read about custom memory management in C++. All these custom new/delete operators, pool allocators usually seem so "geekish" and for "really special purposes only". To tell you the truth, I though I would never need to use them in my life, because standard library allocators "should be enough for everyone". Well, until curious things started to happen...

Well, the first sign of the problems appeared quite long ago. People started to complain that according to system monitor tools (like 'top') Krita ate quite much memory. We could never reproduce it. And what's more 'massif' and internal tile counters always showed we have no memory leaks. We used exactly the number of tiles we needed to store the image of a particular size.

But while making these 8k-image tests, we started to notice that although the number of tiles doesn't grow, the memory reported by 'top' grows quite significantly. Instead of occupying usual 1.3 GiB, which such image would need (layers data + about 400MiB for brushes and textures) reported memory grew up to 3 GiB and higher until OOM Killer woke up and killed Krita. This gave us a clear evidence that we have some problems with fragmentation.

Indeed, during every stoke we have to create about 15000(!) 16KiB objects (tiles). It is quite probable that after a couple of strokes the memory becomes rather fragmented. So we decided to try boost::pool for allocation of these chunks... and it worked! Instead of growing the memory footprint stabilized on 1.3GiB. And that is not counting the fact that boost::pool doesn't free the free'd memory until destruction or explicit purging [0]

Now this new memory management code is already in master! According to some synthetic tests, the painting should become a bit fasted. Not speaking about the much smaller memory usage.

Conclusion:

If you see unusually high memory consumption in your application, and the results measured by massif significantly differ from what you see in 'top', you probably have some fragmentation problem. To proof it, try not to return the memory back to the system, but reuse it. The consumption might fall significantly, especially is you allocate memory in different threads.



[0] - You can release unused memory by explicitly calling release_memory(), but 1) the pool must be ordered, which is worse performance; 2) the release_memory() operation takes about 20-30 seconds(!), so there is no use of it for us.



No comments:

Post a Comment

Followers