BTW, the distribution of unused sizes looks like this:

[% unused]  [total unused size]
    <10%         11.6 KiB
    20%          46.7 KiB
    30%          9.6 KiB
    40%          25.7 KiB
    50%          13.6 KiB
    60%          14.8 KiB
    70%          69.8 KiB
    90%          78.5 KiB

I.e. 78.5 KiB came from buffers that were >=90% unused.

To Herb and/or Ben who play in this space...

I'm currently working on that TODO, but I doubt it will make any difference. For efficiency the images are backed by a slab allocator so it is quite likely that there sill be a great deal of unused data at the end of the slabs at any given point. Your distribution of unused percentages is about what I would expect.

I see. But since we want to shrink most recent allocation, perhaps SkChunkAlloc could introduce 'shrink(N)' method to shrink last allocation by N bytes? That way images should end up tightly packed.

Actually, disregard that, I see that SkChunkAlloc basically allocates one block per image if image is sufficiently large. Shrinking won't help much.

I can see that as of https://skia-review.googlesource.com/c/7112/ SkGlyphCache uses SkArenaAlloc. I'll redo my experiment to see how it helped.

Ping! This bug is assigned and has an owner, but hasn't been updated in over 180 days. Is this still something we want to do?

I suppose this is to dskiba@ if they have time to redo testing with the latest code.

Measurement code is hopelessly obsolete, so I can't redo the measurements. Optimistically marking as fixed.