Crates.io | bonzai |
lib.rs | bonzai |
version | 0.2.1 |
source | src |
created_at | 2018-11-09 02:56:35.651381 |
updated_at | 2018-11-11 05:21:07.334282 |
description | An abstraction for optimizing the memory layout and pointer aliasing of trees. |
homepage | |
repository | https://github.com/gretchenfrage/bonzai.git |
max_upload_size | |
id | 95607 |
size | 79,961 |
Part of Rust's impressive performance comes from Rust storing data in-place by default, which
improves spatial locality, causing the CPU-cache to perform better. Trees, however, often
miss out on this benefit. Trees are recursive structures, so they need to use a form of
indirection. People tend to default to the Box
, since it has the same ownership semantics
as normal, owned data. However, it has the potential to splay data all over the heap, killing
spatial locality.
One technique for resolving this issue is storing a tree's nodes in a contiguous Vec
, and
connecting them with indices. The most direct approach is to implement indexing logic in a way
that is highly coupled with the logic of the tree itself. Unfortunately, this approach bears
a striking resemblance to dealing with raw pointers, both in the bugs it can allow and the
technical complexity is creates.
Bonzai is a safe, zero-cost abstraction over trees which store their nodes in a contiguous
Vec
. Bonzai uses unsafe code to create a safe abstraction which leverages the borrow checker
to verify the validity of operations on trees, while using very few heap allocations.
Send
and Sync
if element isDebug
traitThe bonzai-nbst repository is a rust executable which contains implementations of a naive binary search tree (no balancing), using bonzai in one, and heap allocations in the other. The bonzai-based implementation.
I tested the two implementations with 10,000,000 random tree operations, on a windows 10 laptop. To test it in a real-world scenario with chaotic heap use, I plugged it into a fork of the magog roguelike (which I am not associated with), such that they run in the same process. While this test is very unscientific, bonzai demonstrated a clear performance improvement:
bonzai: 15,786 ms
boxes: 21,338 ms
The Tree
, and nearly all components borrowed from the tree, is generic over two types:
the element type, and a sized array of ChildId
. The element type, T
, is stored for
every node in the tree. Each node may have any combination of children, with any child
index that is a valid index of the ChildId
array.
For example, a binary search tree set of i32
could be represented as a Tree<i32, [ChildId; 2]>
.
A TreeOperation
is a type which borrows mutably from the Tree
, and allows for modification to that
tree. While a TreeOperation
exists, the access to the tree can only be single-threaded. This allows
many operations to mutate the tree with only a immutable reference to the TreeOperation
, directly or
indirectly.
Normally, a node in a tree is owned by its parent. The root node has a special status, where its parent
is considered to be root. However, during the lifetime of a TreeOperation
, a subtree can be detached
from its parent, and it can be considered to be owned by a NodeOwnedGuard
. In this case, the
NodeOwnedGuard
can outlive the guard for its parent. This subtree can be reattached to the main tree
as the child of another node. During the lifetime of the NodeOwnedGuard
, it is possible to mutate it,
as if it were part of the main tree. Alternatively, a NodeOwnedGuard
can be transformed into its
inner element, marking the node and all its children as garbage.
When a NodeOwnedGuard
is dropped, and its subtree is marked as garbage, the elements' destructors will
run sometime between the dropping of the NodeOwnedGuard
and the dropping of the TreeOperation
.
A NodeWriteGuard
is a type which holds mutable access to a subset of the tree (a node and all its
children, recursively). The NodeWriteGuard
can be simultaneously borrowed into a mutable reference
to the element (&mut T
) and a ChildWriteGuard
. The ChildWriteGuard
has the ability to alter the
node's children. Additionally, a NodeWriteGuard
can be turned into a NodeOwnedGuard
, detaching the
guarded subtree.
A NodeWriteGuard
cannot outlive the parent node guard (if the node is root, it cannot
outlive the tree).
A ChildWriteGuard
represents mutable access to a node's children. It can be borrowed from both a
NodeWriteGuard
and a NodeOwnedGuard
. The ChildWriteGuard
can borrow out a NodeWriteGuard
to
its child, or even several different children simultaneously. Additionally, a ChildWriteGuard
can
put an element as a particular child (marking any previous child subtree as garbage), or even attach
an entire detached subtree (a NodeOwnedGuard
) as one of its children.
While a NodeWriteGuard
holds mutable access to a subtree (a node and its children, recursively), a
TreeWriteTraverser
holds mutable access to an entire tree. This allows a TreeWriteTraverser
to
seek the parent node, which a NodeWriteGuard
cannot do. However, this also means that a
TreeWriteTraverser
cannot exist at the same time as any other guards.
Beyond that difference, a TreeWriteTraverser
can be used similarly to a NodeWriteGuard
.
A TreeWriteTraverser
can even be turned or mutably borrowed into a NodeWriteGuard
.
A TreeOperation
and some type of node guard can be conveniently turned into a TreeWriteTraverser
with the traverse_from!
macro.
A NodeReadGuard
is a type which holds immutable access to a subset of the tree. Because a
NodeReadGuard
doesn't mutably borrow anything, it does not need to split into a ChildWriteGuard
and &mut T
. Instead, the NodeReadGuard
immutably dereferences to a T
, and its children
can be accessed directly with a method.
The TreeReadTraverser
is to the NodeReadGuard
as the TreeWriteTraverser
is to the NodeWriteGuard
.
While the NodeReadGuard
holds immutable access to a subset of the tree, a TreeReadTraverser
holds
immutable access to the entire tree. This allows the TreeReadTraverser
to safely traverse to
its parent node.