ShardedQueue
ShardedQueueShardedQueue is currently the fastest collection which can be used
under highest concurrency and load
among most popular solutions, like concurrent-queue -
see benchmarks in directory benches and run them with
bash
cargo bench
bash
cargo add sharded_queue
```rust use std::thread::{availableparallelism}; use shardedqueue::ShardedQueue;
/// How many threads can physically access [ShardedQueue]
/// simultaneously, needed for computing shard_count
let maxconcurrentthreadcount = availableparallelism().unwrap().get();
let shardedqueue = ShardedQueue::new(maxconcurrentthreadcount);
shardedqueue.pushback(1); let item = shardedqueue.popfrontorspinwaititem(); ```
ShardedQueueUnlike in other concurrent queues, FIFO order is not guaranteed.
While it may seem that FIFO order is guaranteed, it is not, because
there can be a situation, when multiple consumers or producers triggered long resize of very large shards,
all but last, then passed enough time for resize to finish, then 1 consumer or producer triggers long resize of
last shard, and all other threads start to consume or produce, and eventually start spinning on
last shard, without guarantee which will acquire spin lock first, so we can't even guarantee that
ShardedQueue::pop_front_or_spin_wait_item will acquire lock before ShardedQueue::push_back on first
attempt
ShardedQueue doesn't track length, since length's increment/decrement logic may change
depending on use case, as well as logic when it goes from 1 to 0 or reverse
(in some cases, like NonBlockingMutex, we don't even add action to queue when count
reaches 1, but run it immediately in same thread), or even negative
(to optimize some hot paths, like in some schedulers,
since it is cheaper to restore count to correct state than to enforce that it can not go negative
in some schedulers)
ShardedQueue doesn't have many features, only necessary methods
ShardedQueue::pop_front_or_spin_wait_item and
ShardedQueue::push_back are implemented
ShardedQueue outperforms other concurrent queues.
See benchmark logic in directory benches and reproduce results by running
bash
cargo bench
| Benchmark name | Operation count per thread | Concurrent thread count | averagetime |
|:-------------------------------------------|---------------------------:|------------------------:|-------------:|
| shardedqueuepushandpopconcurrently | 1000 | 24 | 1.1344 ms |
| concurrentqueuepushandpopconcurrently | 1000 | 24 | 4.8130 ms |
| crossbeamqueuepushandpopconcurrently | 1000 | 24 | 5.3154 ms |
| queuemutexpushandpopconcurrently | 1000 | 24 | 6.4846 ms |
| shardedqueuepushandpopconcurrently | 10000 | 24 | 8.1651 ms |
| concurrentqueuepushandpopconcurrently | 10000 | 24 | 44.660 ms |
| crossbeamqueuepushandpopconcurrently | 10000 | 24 | 49.234 ms |
| queuemutexpushandpopconcurrently | 10000 | 24 | 69.207 ms |
| shardedqueuepushandpopconcurrently | 100000 | 24 | 77.167 ms |
| concurrentqueuepushandpopconcurrently | 100000 | 24 | 445.88 ms |
| crossbeamqueuepushandpopconcurrently | 100000 | 24 | 434.00 ms |
| queuemutexpushandpopconcurrently | 100_000 | 24 | 476.59 ms |
ShardedQueue is designed to be used in some schedulers and NonBlockingMutex
as the most efficient collection under highest
concurrently and load
(concurrent stack can't outperform it, because, unlike queue, which
spreads pop and push contention between front and back,
stack pop-s from back and push-es to back,
which has double the contention over queue, while number of atomic increments
per pop or push is same as in queue)
ShardedQueue uses array of protected by separate Mutex-es queues(shards),
and atomically increments head_index or tail_index when pop or push happens,
and computes shard index for current operation by applying modulo operation to
head_index or tail_index
Modulo operation is optimized, knowing that
x % 2^n == x & (2^n - 1)
, so, as long as count of queues(shards) is a power of two,
we can compute modulo very efficiently using formula
operation_number % shard_count == operation_number & (shard_count - 1)
As long as count of queues(shards) is a power of two and
is greater than or equal to number of CPU-s,
and CPU-s spend ~same time in push/pop (time is ~same,
since it is amortized O(1)),
multiple CPU-s physically can't access same shards
simultaneously and we have best possible performance.
Synchronizing underlying non-concurrent queue costs only
- 1 additional atomic increment per push or pop
(incrementing head_index or tail_index)
- 1 additional compare_and_swap and 1 atomic store
(uncontended Mutex acquire and release)
- 1 cheap bit operation(to get modulo)
- 1 get from queue(shard) list by index
```rust use sharded_queue::ShardedQueue; use std::cell::UnsafeCell; use std::fmt::{Debug, Display, Formatter}; use std::marker::PhantomData; use std::ops::{Deref, DerefMut}; use std::sync::atomic::{AtomicUsize, Ordering};
pub struct NonBlockingMutex<'capturedvariables, State: ?Sized> {
taskcount: AtomicUsize,
taskqueue: ShardedQueue
/// [NonBlockingMutex] is needed to run actions atomically without thread blocking, or context /// switch, or spin lock contention, or rescheduling on some scheduler /// /// Notice that it uses [ShardedQueue] which doesn't guarantee order of retrieval, hence /// [NonBlockingMutex] doesn't guarantee order of execution too, even of already added /// items impl<'capturedvariables, State> NonBlockingMutex<'capturedvariables, State> { pub fn new(maxconcurrentthreadcount: usize, state: State) -> Self { Self { taskcount: AtomicUsize::new(0), taskqueue: ShardedQueue::new(maxconcurrentthreadcount), unsafe_state: UnsafeCell::new(state), } }
/// Please don't forget that order of execution is not guaranteed. Atomicity of operations is guaranteed,
/// but order can be random
pub fn run_if_first_or_schedule_on_first(
&self,
run_with_state: impl FnOnce(MutexGuard<State>) + Send + 'captured_variables,
) {
if self.task_count.fetch_add(1, Ordering::Acquire) != 0 {
self.task_queue.push_back(Box::new(run_with_state));
} else {
// If we acquired first lock, run should be executed immediately and run loop started
run_with_state(unsafe { MutexGuard::new(self) });
/// Note that if [`fetch_sub`] != 1
/// => some thread entered first if block in method
/// => [ShardedQueue::push_back] is guaranteed to be called
/// => [ShardedQueue::pop_front_or_spin_wait_item] will not deadlock while spins until it gets item
///
/// Notice that we run action first, and only then decrement count
/// with releasing(pushing) memory changes, even if it looks otherwise
while self.task_count.fetch_sub(1, Ordering::Release) != 1 {
self.task_queue.pop_front_or_spin_wait_item()(unsafe { MutexGuard::new(self) });
}
}
}
}
/// [Send], [Sync], and [MutexGuard] logic was taken from [std::sync::Mutex]
/// and [std::sync::MutexGuard]
///
/// these are the only places where T: Send matters; all other
/// functionality works fine on a single thread.
unsafe impl<'capturedvariables, State: ?Sized + Send> Send
for NonBlockingMutex<'capturedvariables, State>
{
}
unsafe impl<'capturedvariables, State: ?Sized + Send> Sync
for NonBlockingMutex<'capturedvariables, State>
{
}
/// Code was mostly taken from [std::sync::MutexGuard], it is expected to protect [State] /// from moving out of synchronized loop pub struct MutexGuard< 'capturedvariables, 'nonblockingmutexref, State: ?Sized + 'nonblockingmutex_ref,