Fast persistent immutable data structures
Fast persistent immutable data structures.
Bagwellian hash array-mapped trie, with:
Preliminary benchmarks indicate performance gains of 2–7X for
accrete-based operations (e.g.
associate) compared to equivalent ClojureScript operations (via mori.js).
Preliminary V8 heap profiling indicates space efficiency gains for
HashMap over a range of approximately 0.5–8.0X, depending on collection size and hash distribution, compared to an identical hash map in cljs/mori.
Compared to cljs/mori, heap allocation efficiency for
HashMap may be expected to regress by some factor less than 2X for certain generational cases, namely those yielded over the course of many
The regressive effect described is due to the exclusive definition of a single
HashMapNode type which bears a pairwise array of maximum length 64, as opposed to the definition of two distinct classes,
BitmapIndexedNode for internal nodes, which bears a max-32 array of nodes, and
ArrayNode for terminal nodes, which bears a pairwise max-64 array of key-value pairs.
From this distinction it follows that replication of any given
HashMapNode then redundantly duplicates the bitmap integers associated with any child nodes. As density increases, node replication may thus incur up to twice the memory cost compared to cljs/mori.
It is possible that such regression may be mitigated overall, in part or in whole, as a consequence of the generally shorter trie height that results from storing pairs directly within
This regressive effect is expected, but as yet neither confirmed nor tested.