17 releases (breaking)
Uses new Rust 2021
|0.14.0||Sep 13, 2022|
|0.12.0||Sep 2, 2022|
|0.6.0||Jul 31, 2022|
#73 in Memory management
24 downloads per month
Onsen provides hot Pools for objects. In most cases allocation from such a Pool is faster and offers better locality than the standard allocator. For small to medium sized objects the performance improvement is around 20% or better. For large objects the gains become smaller as caching effects even out. These improvements cover operating on objects because of locality, not just faster allocation speeds.
Onsen pools allocate blocks with exponentially growing sizes. Allocations are served from these blocks. Freed entries are kept in a double linked cyclic freelist. This freelist is kept in weakly ordered and the entry point always point close to where the last action happend to keep the caches hot.
Box, Rc and Sc
Onsen comes with its own
Weak implementations that wrap the underlying
RcPool in a safe way. A
Sc reference counted box without weak reference support is
available as well and provides an advantage for small objects where the weak count would add
For each of these a variant that uses static global pools is avaialble as well.
Allocating from a pool returns
Slot handles. These are lightweight abstractions to memory
addresses, they do not keep a relation to the pool they are allocated from. The rationale for
this design is to make them usable in a VM that uses NaN tagging.
Slots are guarded by typestate policies which prevent some wrong use at compile time.
Slots and Safety
Because of this slots need to be handled with care and certain contracts need to be
enforced. The library provides some help to ensure correctness. Few things can not be asserted
and are guarded by unsafe functions. Higher level API's (Such as
Sc above) can
easily enforce these in a safe way.
- Slots must be given back to the pool they originate from.
- Slots must not outlive the pool they are allocated from.
- When a Pool gets dropped while it still has live allocations it will panic in debug mode.
- When a pool with live allocations gets dropped in release mode it leaks its memory. This is unfortunate but ensures memory safety of the program.
- There is
pool.leak()which drops a pool while leaking its memory blocks. This can be used when one will never try to free memory obtained from that Pool.
- This applies to u64 NaN tags as well.
- Slots must be freed only once.
- This is always asserted. But the assertion may fail when the slot got allocated again.
- Slots are not 'Copy' thus one can not safely free a slot twice but there is an explicit 'copy()' function used by the reference count implementations and the NaN tagging facilities can copy an 'u64' and try to attempt to free this multiple times. These are 'unsafe' functions becasue of that.
- References obtained from slots must not outlive the freeing of the
- This is the main reason that makes the
Slotfreeing functions unsafe. There is no way for a pool to know if references are still in use. One should provide or use a safe abstraction around references to enforce this.
- This is the main reason that makes the
Onsen provides a singlethreaded
Pool, a singlethreaded reference counted
RcPool and a
TPool. Additional features are gated with feature flags.
- parking_lot use parking_lot for the
std::sync::Mutex). This makes sense when parking lot is already in use. There is no significant performance benefit from this in onsen.
- stpool Makes
STPoolavailable, a singlethreaded pool that uses a
ThreadCellwhich is much faster than mutex protected pools. This pools can be moved cooperatively between threads with acquire/release semantics.
- tbox Adds the API for
TScthat use a global pool per type. The advantage is that the box does not need to store a reference to its pool which saves a bit memory and improves locality for small objects.
- st_tbox use
STPoolfor the tbox API, this enables tbox and stpool as well.
st_tbox is the default. This enables the most complete API with best performance.
Onsen pools are optimized for cache locality and with that to some extend for singlethreaded use. It is best to have one pool per type per thread.
TPooladds a mutex to be used in multithreaded cases but its performance is significantly less than the singlethreaded pools but in many cases still better than the std allocator. One will still benefit from locality though.
STPoolis singlethreaded but can be cooperatively passed between threads, its performance is on par with the other singlethreaded pools. This is especially important when one uses
Onsen uses criterion for benchmarking, since onsen is made for singlethreaded application its best to be tested when locked on a single CPU core and lock the core to some frequency well below the max to give more consistent results. At higher priority so it wont be disturbed as much from other programs. On Linux you may do something like:
sudo renice -15 $$ sudo cpupower -c 1 frequenc-set -f 2.8GHz taskset 2 cargo bench