Expand description
Multi-producer multi-consumer lock-free multicast queue1.
Memory-wise all consumers work with the same shared data blocks, so there is no duplication. Queue is unbounded - it grows and shrinks dynamically as needed.
Read performance is stellar - there is only one atomic read per block. Most of the time - this is just plain continuous data reading.
§spmc vs mpmc
Both spmc and mpmc can work in multi-producer mode.
In general, using Arc<spin::Mutex<spmc::Queue<T>>> is more performant
than Arc<mpmc::Queue<T>> from writer perspective.
But! Writing simultaneously from several threads is much faster with mpmc.
The read performance is almost equal, with a slight advantage for spmc.
§Order
spmc is fully ordered.
mpmc ordered within writer’s messages. Which means that all messages from the same Writer will arrive in order.
§Reader memory
All readers have something like Arc for its current block in shared queue. This means that each reader prevents an unread portion of a queue from being dropped.
§target-flags
mpmc use trailing_ones(). So you want to have hardware support for it.
On x86 you need BMI1, there is analog on each cpu architecture.
Or broadcast queue. Each consumer gets every message sent to queue, from the moment of subscription. ↩
Modules§
Structs§
- Cloned
Reader - Cloning queue consumer.
Traits§
- Lending
Reader - Lending queue consumer trait.