Read the documentation.

tinyset

tinyset contains a few collections that are optimized to scale in size well for small numbers of elements, while still scaling well in time (and size) for numbers of elements. We have three set types:

1. Set is basically interchangeable with HashSet, although it does require that its elements implement the Copy trait, since otherwise I would have to learn to write correct unsafe code, which would be scary. It uses FNV hashing when there are large numbers of elements.

2. TinySet is places a stronger requirement on its elements, which must have trait HasInvalid. This is intended for elements that are Copy, are Hash, and have an "invalid" value. For the unsigned integer types, we take their maximum value to mean invalid. This constraint allows us to save a bit more space.

3. Set64 is a set for types that are 64 bits in size or less and are Copy, intended for essentially integer types. This is our most efficient type, since it can store values in less space than std::mem::size_of::<T>(), in the common case that they are small numbers. It is also essentially as fast as any of the other set types (faster than many), and can avoid heap allocations entirely for small sets.

All of these set types will do no heap allocation for small sets of small elements. TinySet will store up to 16 bytes of elements before doing any heap allocation, while Set stores sets up to size 8 without allocation. Set64 will store up to 22 bytes of elements, and if all your elements are small (e.g. 0..22 as u64 it will store them in as few bytes as possible.

All these sets are similar in speed to fnv::HashSet. Set64 is usually faster than fnv::HashSet, sometimes by as much as a factor of 2.

Examples

use tinyset::Set;
let mut s: Set<usize> = Set::new();
s.insert(1);
assert!(s.contains(&1));

use tinyset::TinySet;
let mut s: TinySet<usize> = TinySet::new();
s.insert(1);
assert!(s.contains(&1));

use tinyset::Set64;
let mut s: Set64<usize> = Set64::new();
s.insert(1);
assert!(s.contains(&1));

Hash maps

In addition to the sets that tinyset is named for, we export a couple of space-efficient hash map implentations, which are closely related to Set64 described above. These are

1. Map64 is a map from types that are 64 bits in size or less and are Copy, intended for essentially integer types. The value can be of any type, and the memory use (especially for small or empty maps) is far lower than that of a standard HashMap.
2. Map6464 is a map from types that are 64 bits in size or less and are Copy, to values that are also small and Copy. This is an incredibly space-efficient data type with no heap storage when you have just a few small keys and values. On a 64-bit system, the size of a Map6464 is 48 bytes, and if your keys and values both fit in 8 bits, you can hold 23 items without using the heap. If the keys fit in 16 bits and the values in 8 bits, you can hold 15 itmes without resorting to the heap, and so on. You can even hold a whopping 4 64-bit keys with 8-bit values without resorting to the heap, making this very efficent.

Benchmarks

To run the benchmark suite, cd into bench and then run

cargo run --bin sets --release

This will give you loads of timings and storage requirements for a wide variety of set types.

You can alternatively run

cargo run --bin maps --release

This will give you loads of timings and storage requirements for a variety of map types.

Unfortunately, I don't know an easy way to check the actual memory use for a hashmap, so the benchmarks don't check heap usage. (I used to do this in a fragile way, but cut it.) If you have any suggestions for tracking heap use in a nice way, please let me know!