unordered/doc/modules/ROOT/pages/structures.adoc

[#structures]
= Data Structures

:idprefix: structures_

== Closed-addressing Containers

++++
<style>
  .imageblock > .title {
    text-align: inherit;
  }
</style>
++++

Boost.Unordered sports one of the fastest implementations of closed addressing, also commonly known as https://en.wikipedia.org/wiki/Hash_table#Separate_chaining[separate chaining]. An example figure representing the data structure is below:

[#img-bucket-groups,.text-center]
.A simple bucket group approach
image::bucket-groups.png[align=center]

An array of "buckets" is allocated and each bucket in turn points to its own individual linked list. This makes meeting the standard requirements of bucket iteration straight-forward. Unfortunately, iteration of the entire container is often times slow using this layout as each bucket must be examined for occupancy, yielding a time complexity of `O(bucket_count() + size())` when the standard requires complexity to be `O(size())`.

Canonical standard implementations will wind up looking like the diagram below:

[.text-center]
.The canonical standard approach
image::singly-linked.png[align=center,link=_images/singly-linked.png,window=_blank]

It's worth noting that this approach is only used by pass:[libc++] and pass:[libstdc++]; the MSVC Dinkumware implementation uses a different one. A more detailed analysis of the standard containers can be found http://bannalia.blogspot.com/2013/10/implementation-of-c-unordered.html[here].

This unusually laid out data structure is chosen to make iteration of the entire container efficient by inter-connecting all of the nodes into a singly-linked list. One might also notice that buckets point to the node _before_ the start of the bucket's elements. This is done so that removing elements from the list can be done efficiently without introducing the need for a doubly-linked list. Unfortunately, this data structure introduces a guaranteed extra indirection. For example, to access the first element of a bucket, something like this must be done:

```c++
auto const idx = get_bucket_idx(hash_function(key));
node* p = buckets[idx]; // first load
node* n = p->next; // second load
if (n && is_in_bucket(n, idx)) {
  value_type const& v = *n; // third load
  // ...
}
```

With a simple bucket group layout, this is all that must be done:
```c++
auto const idx = get_bucket_idx(hash_function(key));
node* n = buckets[idx]; // first load
if (n) {
  value_type const& v = *n; // second load
  // ...
}
```

In practice, the extra indirection can have a dramatic performance impact to common operations such as `insert`, `find` and `erase`. But to keep iteration of the container fast, Boost.Unordered introduces a novel data structure, a "bucket group". A bucket group is a fixed-width view of a subsection of the buckets array. It contains a bitmask (a `std::size_t`) which it uses to track occupancy of buckets and contains two pointers so that it can form a doubly-linked list with non-empty groups. An example diagram is below:

[#img-fca-layout]
.The new layout used by Boost
image::fca.png[align=center]

Thus container-wide iteration is turned into traversing the non-empty bucket groups (an operation with constant time complexity) which reduces the time complexity back to `O(size())`. In total, a bucket group is only 4 words in size and it views `sizeof(std::size_t) * CHAR_BIT` buckets meaning that for all common implementations, there's only 4 bits of space overhead per bucket introduced by the bucket groups.

A more detailed description of Boost.Unordered's closed-addressing implementation is
given in an
https://bannalia.blogspot.com/2022/06/advancing-state-of-art-for.html[external article].
For more information on implementation rationale, read the
xref:rationale.adoc#rationale_open_addresing_containers[corresponding section].

== Open-addressing Containers

The diagram shows the basic internal layout of `boost::unordered_flat_set`/`unordered_node_set` and
`boost:unordered_flat_map`/`unordered_node_map`.


[#img-foa-layout]
.Open-addressing layout used by Boost.Unordered.
image::foa.png[align=center]

As with all open-addressing containers, elements (or pointers to the element nodes in the case of
`boost::unordered_node_set` and `boost::unordered_node_map`) are stored directly in the bucket array.
This array is logically divided into 2^_n_^ _groups_ of 15 elements each.
In addition to the bucket array, there is an associated _metadata array_ with 2^_n_^
16-byte words.

[#img-foa-metadata]
.Breakdown of a metadata word.
image::foa-metadata.png[align=center]

A metadata word is divided into 15 _h_~_i_~ bytes (one for each associated
bucket), and an _overflow byte_ (_ofw_ in the diagram). The value of _h_~_i_~ is:

  - 0 if the corresponding bucket is empty.
  - 1 to encode a special empty bucket called a _sentinel_, which is used internally to
  stop iteration when the container has been fully traversed.
  - If the bucket is occupied, a _reduced hash value_ obtained from the hash value of
  the element.

When looking for an element with hash value _h_, SIMD technologies such as
https://en.wikipedia.org/wiki/SSE2[SSE2] and
https://en.wikipedia.org/wiki/ARM_architecture_family#Advanced_SIMD_(Neon)[Neon] allow us
to very quickly inspect the full metadata word and look for the reduced value of _h_ among all the
15 buckets with just a handful of CPU instructions: non-matching buckets can be
readily discarded, and those whose reduced hash value matches need be inspected via full
comparison with the corresponding element. If the looked-for element is not present,
the overflow byte is inspected:

- If the bit in the position _h_ mod 8 is zero, lookup terminates (and the
element is not present).
- If the bit is set to 1 (the group has been _overflowed_), further groups are
checked using https://en.wikipedia.org/wiki/Quadratic_probing[_quadratic probing_], and
the process is repeated.

Insertion is algorithmically similar: empty buckets are located using SIMD,
and when going past a full group its corresponding overflow bit is set to 1.

In architectures without SIMD support, the logical layout stays the same, but the metadata
word is codified using a technique we call _bit interleaving_: this layout allows us
to emulate SIMD with reasonably good performance using only standard arithmetic and
logical operations.

[#img-foa-metadata-interleaving]
.Bit-interleaved metadata word.
image::foa-metadata-interleaving.png[align=center]

A more detailed description of Boost.Unordered's open-addressing implementation is
given in an
https://bannalia.blogspot.com/2022/11/inside-boostunorderedflatmap.html[external article].
For more information on implementation rationale, read the
xref:#rationale_open_addresing_containers[corresponding section].

== Concurrent Containers

`boost::concurrent_flat_set`/`boost::concurrent_node_set` and
`boost::concurrent_flat_map`/`boost::concurrent_node_map` use the basic
xref:#structures_open_addressing_containers[open-addressing layout] described above
augmented with synchronization mechanisms.


[#img-cfoa-layout]
.Concurrent open-addressing layout used by Boost.Unordered.
image::cfoa.png[align=center]

Two levels of synchronization are used:

* Container level: A read-write mutex is used to control access from any operation
to the container. Typically, such access is in read mode (that is, concurrent) even
for modifying operations, so for most practical purposes there is no thread
contention at this level. Access is only in write mode (blocking) when rehashing or
performing container-wide operations such as swapping or assignment.
* Group level: Each 15-slot group is equipped with an 8-byte word containing:
  ** A read-write spinlock for synchronized access to any element in the group.
  ** An atomic _insertion counter_ used for optimistic insertion as described
  below.

By using atomic operations to access the group metadata, lookup is (group-level)
lock-free up to the point where an actual comparison needs to be done with an element
that has been previously SIMD-matched: only then is the group's spinlock used.

Insertion uses the following _optimistic algorithm_:

* The value of the insertion counter for the initial group in the probe
sequence is locally recorded (let's call this value `c0`).
* Lookup is as described above. If lookup finds no equivalent element,
search for an available slot for insertion successively locks/unlocks
each group in the probing sequence.
* When an available slot is located, it is preemptively occupied (its
reduced hash value is set) and the insertion counter is atomically
incremented: if no other thread has incremented the counter during the
whole operation (which is checked by comparing with `c0`), then we're
good to go and complete the insertion, otherwise we roll back and start
over.

This algorithm has very low contention both at the lookup and actual
insertion phases in exchange for the possibility that computations have
to be started over if some other thread interferes in the process by
performing a succesful insertion beginning at the same group. In
practice, the start-over frequency is extremely small, measured in the range
of parts per million for some of our benchmarks.

For more information on implementation rationale, read the
xref:#rationale_concurrent_containers[corresponding section].