Skip to content

VecDeque-like fast, unbounded, mpmc/spmc concurent FIFO message queue. Lockless reads, write-lock writes.

License

Apache-2.0, MIT licenses found

Licenses found

Apache-2.0
LICENSE-APACHE
MIT
LICENSE-MIT
Notifications You must be signed in to change notification settings

tower120/rc_event_queue

Repository files navigation

crates.io Docs CI

Warning

Project is deprecated in favor of chute.

Chute is a continuation of this project, featuring truly lock-free MPMC writers that are superlinearly faster in highly concurrent scenarios.

Reader counted event queue

Fast, concurrent FIFO event queue (or message queue). Multiple consumers receive every message.

  • mpmc (multi-producer multi-consumer) - lock-free read, locked write.
  • spmc (single-producer multi-consumer) - lock-free read, lock-free write.

Write operations never block read operations. Performance consumer oriented. Mostly contiguous memory layout. Memory consumption does not grow with readers number.

Performance

Have VERY low CPU + memory overhead. Most of the time reader just do 1 atomic load per iter() call. That's all!

Single-threaded.

Read - close to VecDeque! Write:

  • mpmc - push 2x slower then VecDeque. extend with at least 4 items, close to VecDeque.
  • spmc - equal to VecDeque!

Multi-threaded.

Read - per thread performance degrades slowly, with each additional simultaneously reading thread. (Also remember, since rc_event_queue is message queue, and each reader read ALL queue - adding more readers does not consume queue faster)

Write - per thread performance degrades almost linearly, with each additional simultaneously writing thread. (Due to being locked). Not applicable to spmc.

N.B. But if there is no heavy contention - performance very close to single-threaded case.

See mpmc benchmarks.

Principle of operation

See doc/principle-of-operation.md.

Short version - EventQueue operates on the chunk basis. EventQueue does not touch EventReaders . EventReaders always "pull" from EventQueue. The only way EventReader interact with EventQueue - by increasing read counter when switching to next chunk during traverse.

Usage

API doc

use rc_event_queue::prelude::*;
use rc_event_queue::mpmc::{EventQueue, EventReader};

let event = EventQueue::<usize>::new();
let mut reader1 = EventReader::new(event);
let mut reader2 = EventReader::new(event);

event.push(1);
event.push(10);
event.push(100);
event.push(1000);

fn sum (mut iter: impl LendingIterator<ItemValue = usize>) -> usize {
    let mut sum = 0;
    while let Some(item) = iter.next() {
        sum += item;
    }
    sum
}

assert!(sum(reader1.iter()) == 1111);
assert!(sum(reader1.iter()) == 0);
assert!(sum(reader2.iter()) == 1111);
assert!(sum(reader2.iter()) == 0);

event.extend(0..10);
assert!(sum(reader1.iter()) == 55);
assert!(sum(reader2.iter()) == 55);

clear:

event.push(1);
event.push(10);
event.clear();
event.push(100);
event.push(1000);

assert!(sum(reader1.iter()) == 1100);
assert!(sum(reader2.iter()) == 1100);

clear/truncate_front have peculiarities - chunks occupied by readers, will not be freed immediately.

Emergency cut

If any of the readers did not read for a long time - it can retain queue from cleanup. This means that queue capacity will grow. On long runs with unpredictable systems, you may want to periodically check total_capacity, and if it grows too much - you may want to force-cut/clear it.

if event.total_capacity() > 100000{
    // This will not free chunks occupied by readers, but will free the rest.
    // This should be enough, to prevent memory leak, in case if some readers
    // stop consume unexpectedly.
    event.truncate_front(1000);     // leave some of the latest messages to read
    
    // If you set to Settings::MAX_CHUNK_SIZE to high value,
    // this will reduce chunk size.
    event.change_chunk_size(2048);

    // If you DO have access to all readers (you probably don't) - 
    // this will move readers forward, and free the chunks occupied by readers.
    // Under normal conditions, this is not necessary, since readers will jump
    // forward to another chunk immediately on the next iter() call.
    for reader in readers{
        reader.update_position();
        // reader.iter();   // this have same effect as above
    }
}

Even if some reader will stop read forever - you'll only lose/leak chunk directly occupied by reader.

Optimisation

CLEANUP

Set CLEANUP to Never in Settings, in order to postpone chunks deallocations.

use rc_event_reader::mpmc::{EventQueue, EventReader, Settings};

struct S{} impl Settings for S{
    const MIN_CHUNK_SIZE: u32 = 4;
    const MAX_CHUNK_SIZE: u32 = 4096;
    const CLEANUP: CleanupMode = CleanupMode::Never;
}

let event = EventQueue::<usize, S>::new();
let mut reader = event.subscribe();

event.extend(0..10);
sum(reader.iter()); // With CLEANUP != Never, this would cause chunk deallocation

...

event.cleanup();   // Free used chunks

double_buffering

Use double_buffering feature. This will reuse biggest freed chunk. When EventQueue reach its optimal size - chunks will be just swapped, without alloc/dealloc.

Soundness

EventQueue covered with tests. Miri tests. Loom tests. See doc/tests.md

About

VecDeque-like fast, unbounded, mpmc/spmc concurent FIFO message queue. Lockless reads, write-lock writes.

Topics

Resources

License

Apache-2.0, MIT licenses found

Licenses found

Apache-2.0
LICENSE-APACHE
MIT
LICENSE-MIT

Stars

Watchers

Forks

Packages

No packages published

Languages