St³ — the Stealing Static Stack

A very fast lock-free, bounded, work-stealing queue with stack-like (LIFO) semantic for the worker thread and FIFO stealing.

Overview

The Go scheduler and the Tokio runtime are examples of high-performance schedulers that rely on fixed-capacity (bounded) work-stealing queues to avoid the allocation and synchronization overhead associated with unbounded queues such as the Chase-Lev work-stealing deque (used in particular by Crossbeam Deque). This is a natural design choice for schedulers that use a global injector queue as the latter often can, at nearly no extra cost, buffer the overflow should a local queue become full.

The Go and Tokio schedulers use FIFO local queues to ensure fairness. For schedulers that do not require fairness, however, LIFO queues can provide better performance since dequeued tasks are more likely to have their associated data still available in the CPU cache[^1]. Moreover, contention between workers and stealers is reduced since items are popped and stolen from opposite ends.

For such use-cases, St³ constitutes a faster, fixed-size alternative to the Chase-Lev deque with even slightly better performance than the Tokio work-stealing queue.

[^1]: Go and Tokio actually both use a single-task LIFO slot that bypasses the FIFO queue for this very reason.

Usage

Add this to your Cargo.toml:

[dependencies]
st3 = "0.3.1"

Example

use std::thread;
use st3::{Worker, B256};

// Push 4 items into a queue of capacity 256.
let worker = Worker::<_, B256>::new();
worker.push("a").unwrap();
worker.push("b").unwrap();
worker.push("c").unwrap();
worker.push("d").unwrap();

// Steal items concurrently.
let stealer = worker.stealer();
let th = thread::spawn(move || {
    let other_worker = Worker::<_, B256>::new();

    // Try to steal half the items and return the actual count of stolen items.
    match stealer.steal(&other_worker, |n| n/2) {
        Ok(actual) => actual,
        Err(_) => 0,
    }
});

// Pop items concurrently.
let mut pop_count = 0;
while worker.pop().is_some() {
    pop_count += 1;
}

// Does it add up?
let steal_count = th.join().unwrap();
assert_eq!(pop_count + steal_count, 4);

Safety — a word of caution

This is a low-level primitive and as such its implementation relies on unsafe. The test suite makes extensive use of Loom to assess its correctness. As amazing as it is, however, Loom is only a tool: it cannot formally prove the absence of data races.

Before St³ sees wider use in the field and receives greater scrutiny, you should exercise caution before using it in mission-critical software. It is a new concurrent algorithm and it is therefore possible that soundness issues will be discovered that weren't caught by the test suite.

Performance

St³ uses no atomic fences and very few atomic Read-Modify-Write (RMW) operations. Similarly to the Tokio queue, it needs no RMW for push and only one for pop. Stealing operations require only a single RMW (one less than the Tokio queue).

The first benchmark measures performance in the single-threaded, no-stealing case: a series of 64 push operations (or 256 in the large-batch case) is followed by as many pop operations.

Test CPU: i5-7200U.

The second benchmark is a synthetic test that aims at characterizing multi-threaded performance with concurrent stealing. It uses a toy work-stealing executor which schedules on each of 4 workers an arbitrary number of tasks (from 1 to 256), each task being repeated by re-injection onto its worker an arbitrary number of times (from 1 to 100). The number of tasks initially assigned to each workers and the number of times each task is to be repeated are deterministically pre-determined with a pseudo-RNG, meaning that the workload is the same for all benchmarked queues. All queues use the Crossbeam Dequeue work-stealing strategy: half of the tasks are stolen, up to a maximum of 32 tasks. Nevertheless, the re-distribution of tasks via work-stealing is ultimately non-deterministic as it is affected by thread timing.

Given the somewhat simplistic and subjective design of the benchmark, the figures below must be taken with a grain of salt. That being said, the relative performance of the different queues appear to be qualitatively unaffected by the specific values of the benchmark parameters.

Test CPU: i5-7200U.

(Mis)Features

Just like the Tokio queue, St³ is susceptible to ABA. For instance, in a naive implementation, if a steal operation was preempted at the wrong moment for exactly the time necessary to pop a number of items equal to the queue capacity while pushing less items than are popped, once resumed the stealer could attempt to steal more items than are available. ABA is overcome by using buffer positions that can index many times the actual buffer capacity so as to increase the cycle period beyond worst-case preemption.

Unlike Tokio, St³ will by default use 32-bit rather than 16-bit buffer positions. Relying on 16 bits to prevent ABA is fairly risky whereas 32 bits should in practice provides full resilience. This requires the use of 64-bit atomics, however, which on 32-bit targets may not be supported (e.g. MIPS) or may be slower (e.g. ARMv7). If you like to live dangerously, you can still elect to disable the default long_counter feature and use 16-bit buffer positions instead:

[dependencies]
st3 = { version = "0.3.1", default-features = false }

Note that disabling this feature has no effect on 64-bit targets: those will still use 32-bit buffer positions.

Acknowledgements

Although the implementation ended up quite different, the Tokio queue was an inspiration which also helped set the goal in terms of performance.

Tokio's queue is itself a modified version of Go's work-stealing queue. Go uses something akin to a Seqlock pattern where stealers optimistically read all items marked for stealing and later discard them if they have been concurrently evicted from the queue. Because of its stricter aliasing rules, Rust makes this pattern hard to implement so the Tokio queue was designed with the ability to "book" the items beforehand, an idea which St³ borrowed.

License

This software is licensed under the Apache License, Version 2.0 or the MIT license, at your option.

Some assets of the test suite and benchmark may be licensed under different terms, which are explicitly outlined within those assets.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.