A library that implements smart batching between a producer and a consumer. In other words, a single-producer single-consumer queue that tries to ensure that the either side runs for a while (to make use of L1 caches) before hibernating again.
One nice feature of this queue is that all buckets are preallocated, so unless you use the recv_batch
or try_recv_batch
methods no allocations are done as part of sending and receiving items.
# tokio_test::block_on(async {
use batch_queue::{batch_queue, pipe};
let (tx, mut rx) = batch_queue::<u32, 128>(20);
const N: u32 = 10_000_000;
tokio::spawn(async move {
let stream = futures::stream::iter(0..N);
pipe(stream, tx).await;
});
let mut x = 0;
while let Ok(iter) = rx.recv().await {
for y in iter {
assert_eq!(y, x);
x += 1;
}
}
assert_eq!(x, N);
# })
Here, the iterator returned by recv()
allows you to consume the bucket contents without allocating.
On my AMD Hetzner box it takes about 6ns per item.
The queue is modeled as a boxed slice of buckets, which each is an appropriately sized array of MaybeUninit
-ialized items and a fill level.
Reader and writer maintain a current position each, which includes the bucket index as well as a cycle counter (to detect when the writer is exactly one full round ahead of the reader).
Inserting an item:
- check whether the writer’s bucket is available for writing (i.e. it is not the reader’s bucket in the previous cycle)
- if so, insert into next slot in bucket; if full, move write bucket position
Declaring end of current batch:
- check whether current write bucket has elements ⇒ move write bucket position
Taking a bucket:
- check whether the writer’s bucket is ahead of the reader’s bucket ⇒ take it and move read bucket position