- Summary
- Why Another Bump Allocator?
- Example
- Flexibility
- Comparison with
bumpalo - Crate Features
- UTF-16 Strings
- Building DSTs
- How it Works
- Testing and Validation
Fast and flexible arena-based bump allocator.
multitude is an arena-based bump allocator designed to improve the performance of applications that have phase-oriented logic, which
is when groups of related allocations live and die together. Service request handling and parsers are two examples of this pattern which usually
benefit from a bump allocator.
multitude works by accumulating large chunks of memory allocated from the system and then carving out smaller pieces of it for application use
using a fast bump allocation strategy, which is considerably faster than allocating from the system. The downside however is that the individual allocations
can't be freed separately. Instead, memory is reclaimed and returned to the system in bulk when the entire arena is dropped.
The Rust ecosystem has a few bump allocators, the most popular being bumpalo.
multitude uses a different implementation strategy and has a richer API surface making it suitable for more
use cases. The main features that set multitude apart are:
-
Flexibility.
multitudeprovides multiple allocation styles, all of which can coexist in the same arena:- Mutable references with lifetimes tied to the arena (
&mut T,&mut str,&mut [T]). - Reference-counted smart pointers ([
Rc], [RcStr]) for single-threaded sharing. - Atomic reference-counted smart pointers ([
Arc], [ArcStr]) for cross-thread sharing. - Owned, mutable smart pointers ([
Box], [BoxStr]).
- Mutable references with lifetimes tied to the arena (
-
Early Reclamation. In many situations,
multitudecan reclaim memory from individual chunks as soon as their reference counts drop to zero, without waiting for the entire arena to be dropped. This allows for more efficient memory usage in long-running arenas with many short-lived allocations. -
Smart Pointers Can Outlive the Arena. The reference-counted smart pointers produced by
multitudecan keep their owning chunk alive even after the arena itself has been dropped, allowing for more flexible memory management and longer-lived data structures. -
Drop Support.
multitudeautomatically runsDropfor allocated values at the appropriate time. -
Efficient Immutable String References.
multitudeprovides [RcStr], [ArcStr], and [BoxStr] — single-pointer (8 bytes) smart pointers to UTF-8 strings stored in the arena. Refcounted, atomic-refcounted, and owned-mutable variants respectively, all sharing the same compact layout. -
Efficient Mutable Strings and Vectors. With the
buildersCargo feature (default-on),multitudeprovidesStringandVecwhich are growable collections that live in the arena and can be frozen into compact reference-counted smart pointers when you're done building. -
Dynamically-Sized Types.
multitudesupports dynamically-sized types (DSTs) like slices and strings, allowing you to allocate and manage them in the arena with the same flexibility as sized types. Thedst-factorycrate is a great companion for building DSTs in the arena. -
format!-style Macro.multitudeincludes aformat!-style macro that allows you to create formatted strings directly in the arena, avoiding intermediate allocations and copies. -
UTF-16 Support. With the
utf16Cargo feature,multitudeprovides a parallel set of arena-resident UTF-16 string types ([RcUtf16Str], [ArcUtf16Str], [BoxUtf16Str],Utf16String) and aformat_utf16!macro for FFI / Windows / JS-engine interop without per-call transcoding at every boundary. -
#![no_std]Support.multitudecan be used in#![no_std]environments, making it suitable for embedded systems and other resource-constrained contexts.
use multitude::Arena;
let arena = Arena::new();
// Cheap reference-counted allocation of any user type.
struct Point { x: f64, y: f64 }
let p = arena.alloc_rc(Point { x: 3.0, y: 4.0 });
let p2 = p.clone();
assert_eq!(p.x, p2.x);
// Single-pointer immutable strings.
let name = arena.alloc_str_rc("Alice");
assert_eq!(&*name, "Alice");
// format! macro returning an String (call .into_arena_str() to
// freeze into a compact 8-byte RcStr).
let greeting = multitude::builders::format!(in &arena, "Hello, {}!", "world");
assert_eq!(&*greeting, "Hello, world!");multitude supports a variety of ways to allocate data and track it over time.
The simplest use of the arena is to get plain mutable references. The lifetime of those references is then tied to the arena's own lifetime.
let arena = multitude::Arena::new();
let x: &mut u32 = arena.alloc(42);
let y: &mut u32 = arena.alloc(100);
*x += 1;
*y += 1;
assert_eq!(*x, 43);
assert_eq!(*y, 101);
// Strings and slices too:
let s: &mut str = arena.alloc_str("hello");
let v: &mut [i32] = arena.alloc_slice_copy(&[1, 2, 3]);These references can't outlive the arena, which limits their use. But they are the fastest and most efficient way to allocate from the arena, so if the lifetime constraints are tolerable, simple references are the way to go.
Smart pointers ([Rc], [Arc], [Box] and their str variations) work in a way similar to the like-named types
in the standard library, except that they reference addresses within an arena.
use multitude::Rc;
struct Point { x: f64, y: f64 }
let p: Rc<Point> = {
let arena = multitude::Arena::new();
arena.alloc_rc(Point { x: 3.0, y: 4.0 })
// arena dropped here
};
assert_eq!(p.x, 3.0);Although [Arena] itself is single-threaded (!Send and !Sync), the arc-family of types (e.g., [Arc]) enable cross-thread sharing.
let arena = multitude::Arena::new();
let shared = arena.alloc_arc(42_u64);
let h = std::thread::spawn(move || *shared);
assert_eq!(42, h.join().unwrap());[Box] is a unique owner whose Drop runs T::drop immediately
when the smart pointer is dropped and provides &mut T access, similar to
[alloc::boxed::Box] but backed by the arena.
let arena = multitude::Arena::new();
let mut v = arena.alloc_box(vec![1, 2, 3]);
v.push(4);
assert_eq!(*v, vec![1, 2, 3, 4]);
drop(v); // Vec's drop runs here, freeing its heap buffer.Vec and String are growable collections that live in
the arena.
Additionally, you can use an arena as
the allocator for any type from the [allocator_api2] ecosystem
(including hashbrown::HashMap).
use multitude::Arena;
use multitude::builders::{Vec, CollectIn};
let arena = Arena::new();
let mut v = arena.alloc_vec::<i32>();
for i in 0..5 { v.push(i); }
// CollectIn trait for iterator collection.
let squares: Vec<i32, _> = (1..=5).map(|i| i * i).collect_in(&arena);
assert_eq!(squares.as_slice(), &[1, 4, 9, 16, 25]);String and Vec are designed as transient
builders. They carry a data pointer + length + capacity + arena reference.
Once you're done building, you can freeze them into immutable smart pointers:
String::into_arena_str→ [RcStr] (8 bytes). The freeze itself is O(1) — no copy, no new allocation.Vec::into_arena_rc→Rc<[T]>(16-byte slice fat pointer; immutable, cloneable, refcount-based). ForT: !Drop, the freeze is O(1) too.
Both freezes also reclaim any unused capacity left in the buffer when the conditions allow it, so those bytes become available for the next allocation.
use multitude::{Arena, RcStr};
let arena = Arena::new();
// Build phase: 32-byte builder, alive briefly.
let mut builder = arena.alloc_string();
builder.push_str("hello, ");
builder.push_str("world");
// Freeze for storage: 8-byte single-pointer smart pointer. O(1) — no copy.
let stored: RcStr = builder.into_arena_str();
assert_eq!(&*stored, "hello, world");Use this pattern whenever you'd be storing many strings or slices long-term — the per-pointer savings (16 bytes for strings, 8 for slices) add up quickly across millions of items, and the frozen smart pointers are also cheaper to clone.
bumpalo is the closest crate in
spirit; here's how multitude differs.
| Capability | bumpalo |
multitude |
|---|---|---|
| Bump allocation | ✅ | ✅ |
Simple references (&'arena mut T) |
✅ Bump::alloc |
✅ [Arena::alloc] |
Allocator trait integration |
✅ via allocator-api2 |
✅ via allocator-api2 |
| Reclamation granularity | Whole arena at reset | Per chunk, as refcounts hit 0 (refcount smart pointers); whole-arena (simple references) |
| Smart pointers | ❌ (raw &'bump T) |
✅ [Rc], [Arc], [RcStr] |
| Smart pointers outlive the arena | ❌ | ✅ ([Rc] / [Arc] / [Box] and their str variants — simple references are lifetime-bound) |
| Cross-thread sharing of individual values | ❌ | ✅ via [Arc] |
Automatic per-object Drop |
Only via bumpalo::boxed::Box |
✅ Automatic (refcount smart pointers drop at chunk teardown; [Box] / [BoxStr] drop at smart pointer drop; simple references drop at arena drop) |
Owned single smart pointer (Drop on drop) |
bumpalo::boxed::Box |
[Box] |
| Single-pointer string smart pointers | ❌ (&str is 16 bytes) |
✅ [RcStr] / [ArcStr] / [BoxStr] are 8 bytes |
| Growable collections | ✅ bumpalo::collections |
✅ Vec, String |
format!-style macro |
✅ | ✅ |
| UTF-16 strings | ❌ | ✅ via RcUtf16Str / ArcUtf16Str / BoxUtf16Str / Utf16String (gated on the utf16 feature) |
Dynamically-sized types (DSTs, e.g. dyn Trait, [T]) |
❌ | ✅ via the dst module (gated on the dst feature) |
zerocopy integration |
❌ | ✅ Zero-initialized allocations for FromZeros types (gated on the zerocopy feature) |
bytemuck integration |
❌ | ✅ Zero-initialized allocations for Zeroable types (gated on the bytemuck feature) |
bytesbuf integration |
❌ | ✅ Arena implements bytesbuf::mem::Memory for arena-backed byte buffers (gated on the bytesbuf feature) |
#![no_std] |
✅ | ✅ |
std— enablesstd::io::Write-style integration where applicable. Disable for#![no_std]environments (the crate still requiresalloc).builders— enables the [builders] module, which contains the growable collection types (String,Vec,Utf16String), theCollectIn/FromIteratorIntraits, and theformat!/vec!/format_utf16!macros. Also adds the matchingArena::alloc_string/alloc_vec/alloc_utf16_stringmethods. Disable for the leanest builds — the smart pointers ([RcStr], [Rc], etc.) andArena::alloc_str_*/alloc_*_slice_*still work without it.stats— enables runtime instrumentation counters returned byArena::stats. Disable for the tightest allocation throughput when you don't need observability.serde— addsSerializeimpls for [RcStr], [ArcStr], and (withbuilders)StringandVec. Withserde + utf16, also addsSerializeimpls for the UTF-16 types (transcoded to UTF-8 on the wire).dst— enables thedstmodule for constructing true dynamically-sized types and trait objects in the arena via [Arena::alloc_dst_arc] / [Arena::alloc_dst_rc] / [Arena::alloc_dst_box], plus eightArena::alloc_slice_*_boxmethods.utf16— adds a parallel UTF-16 string surface (RcUtf16Str,ArcUtf16Str,BoxUtf16Str, plusUtf16Stringandformat_utf16!when combined withbuilders) backed by thewidestringcrate. All length and capacity values are counted inu16elements (matchingwidestring::Utf16Str::len()).zerocopy— providesZerocopyViewfor safe zero-initialized allocation of types implementingzerocopy::FromZeros. Access via [Arena::zerocopy()].bytemuck— providesBytemuckViewfor safe zero-initialized allocation of types implementingbytemuck::Zeroable. Access via [Arena::bytemuck()].bytes— adds [From] conversions fromArc<[u8]>and [ArcStr] intobytes::Bytes, enabling zero-copy integration with the Tokio / Hyper async ecosystem.bytesbuf— implementsbytesbuf::mem::Memorydirectly on [Arena], so thatBytesBufbuffers can be backed by arena chunks. Impliesstd.
The default-enabled features are std and builders.
With the utf16 Cargo feature enabled, multitude exposes a
parallel set of arena-resident UTF-16 string types that mirror the
UTF-8 surface element-for-element:
| UTF-8 type | UTF-16 analogue |
|---|---|
[RcStr] |
RcUtf16Str |
[ArcStr] |
ArcUtf16Str |
[BoxStr] |
BoxUtf16Str |
String |
Utf16String |
Strict (validated) UTF-16 only — lone surrogates are rejected. The
types are interoperable with widestring::Utf16Str /
widestring::Utf16String for I/O and FFI bridging.
use multitude::Arena;
use widestring::utf16str;
let arena = Arena::new();
// From a validated &Utf16Str literal:
let s = arena.alloc_utf16_str_rc(utf16str!("hello, world"));
assert_eq!(&*s, utf16str!("hello, world"));
// Or transcode from a &str:
let s2 = arena.alloc_utf16_str_rc_from_str("hello");
assert_eq!(&*s2, utf16str!("hello"));
// Build incrementally and freeze:
let mut b = arena.alloc_utf16_string();
b.push_str(utf16str!("abc"));
b.push_from_str("123");
let frozen = b.into_arena_utf16_str();
assert_eq!(&*frozen, utf16str!("abc123"));
// format!-style:
let name = "Alice";
let greeting = multitude::builders::format_utf16!(in &arena, "Hello, {name}!");
assert_eq!(greeting.as_utf16_str(), utf16str!("Hello, Alice!"));With the dst Cargo feature enabled, [Arena] exposes
[Arena::alloc_dst_arc] / [Arena::alloc_dst_rc] /
[Arena::alloc_dst_box] (and their try_* siblings) for
constructing values whose layout is only known at runtime (custom
DSTs, fat pointers, trait objects).
Each of these takes a Layout, a
pointer-metadata value (e.g. a slice length, a DynMetadata), and
a closure that initializes the buffer through a typed fat pointer.
For most users, the dst-factory companion crate is the
recommended high-level driver; the low-level interface looks like:
use core::alloc::Layout;
use multitude::Arena;
let arena = Arena::new();
// Allocate a 5-byte slice in the arena as a `Box<[u8]>`.
let layout = Layout::array::<u8>(5).unwrap();
let b: multitude::Box<[u8]> = unsafe {
arena.alloc_dst_box::<[u8]>(layout, 5, |fat: *mut [u8]| {
let p = fat.cast::<u8>();
for i in 0..5 { p.add(i).write(i as u8); }
})
};
assert_eq!(&*b, &[0, 1, 2, 3, 4]);The same feature also enables eight Arena::alloc_slice_*_box
methods that produce Box<[T]> directly (mirroring the
existing _rc/_arc slice methods).
A high-level sketch of what happens when you use the crate. For the gory implementation details, read the source.
The arena owns a list of large memory chunks allocated from the system. Each user allocation carves out a piece of the current chunk via a fast bump-pointer increment. When the current chunk can't fit the next request, the arena rotates to a fresh chunk. Allocations larger than a chunk's usable capacity get their own dedicated oversized chunks.
The smart-pointer types ([Rc], [Arc], [Box])
are deliberately compact — typically one pointer plus zero-sized
markers. Reference counts live with the chunk, not with each smart
pointer, so cloning and dropping cost a single counter update.
The single-pointer string types ([RcStr], [ArcStr],
[BoxStr]) go further: 8 bytes per smart pointer (vs 16
for &str).
When the last smart pointer into a chunk is dropped, the arena runs every destructor for values it hosted, then either parks the chunk in a free-list cache for fast reuse or returns its memory to the system allocator. This per-chunk reclamation lets long-lived arenas with churn keep their memory footprint bounded.
Smart pointers ([Rc] / [Arc] / [Box] and their
str variants) may outlive the arena: their chunks stay alive
as long as any smart pointer references them, even after the arena
itself has been dropped.
References returned by [Arena::alloc] don't carry refcounts;
their chunks live until arena drop. The borrow checker keeps them
lifetime-bound to the arena.
The arena itself is !Send and !Sync. Cross-thread sharing
happens at the smart-pointer level via [Arc] /
[ArcStr], which use atomic refcounts. Single-threaded code
that uses [Rc] only stays on cheap non-atomic operations.
Since mutltitude contains a lot of tricky unsafe code, much effort went to ensure it is correct:
-
Clippy. Catches common mistakes and enforces idiomatic Rust patterns through hundreds of additional lints.
-
Miri. Detects undefined behavior and memory safety violations by executing code in an interpreted environment.
-
cargo-careful. Runs tests with the standard library's internal safety checks enabled to catch violated invariants. -
cargo-mutants. Validates test suite depth by injecting synthetic bugs to ensure your tests can actually detect them. -
loom. Exhaustively tests concurrent code by exploring all possible thread interleavings to find subtle race conditions. -
bolero. Combines fuzzing and property-based testing to automatically surface edge cases with randomized input generation.
For performance, we use complementary criterion and gungraun benchmarks to get wall-clock time and instruction-level metrics.