Lately, I got a large speed-up using a N-heap rather than a binary heap in a priority queue. But it's not the only binary trees I had, Phobos' Red-Black tree is like that.
Could we stuff more items per tree node?
As memory gets relatively more expensive than CPU, does that wins some precious cycles?
Looking at the Map I was using, I've compared a few different equivalent for string[int] / Map!(int, string) / associative arrays.
- one is the regular builtin AA, a hashmap (druntime)
- one is based upon Phobos red-black tree (rbtree.d), adapted to use manual allocation
- one is based on a new Boost-licensed, B-tree implementation in Dplug (btree.d)
- one is the TreeMap in emsi-container package, itself using a T-Tree (treemap.d, ttree.d)
The results are summed up in this presentation: https://dplug.org/tutorials/Dplug%20Tutorials%2018%20-%20The%20Case%20Against%20Binary%20Trees.pdf
Notes:
- For an in-memory collection, a B-Tree seems generally more performant than a Red-Black tree nowadays. In fact I believe any tree with more than 2 children will outperform a tree with only 2 children per node.
- However Red-Black Tree are used because C++ std::map must be able to insert without invalidating existing iterators. It's possible with RB-Tree but not B-Tree.
- In my benchmarks I don't understand why
emsi_containers
T-Tree act like this at look-up; it looks slower than one could expect. Bug? So I wasn't able to rightfully compare T-Tree with B-Tree. It shouldn't be the GC, but possibly it's adding the GC roots that do this? - builtin AA uses a lot of tricks in the book and, as a hashmap, can outperform the others who are instead sorted collections. But not necessarily at look-up.