Multithreading is not enabled by default when using subcrates, leading to confusing results #21585
Description
When spawning 8 blocking tasks (each with thread::sleep(10ms)) via both bevy_tasks::ComputeTaskPool::get().scope and ECS schedule (MultiThreaded executor), actual runtime is 80ms+ instead of the expected ~10ms. This means all tasks are run strictly serially, with no true thread pool concurrency.
bevy_app bevy_task 0.17.2 default features
[Optional] Relevant system information
Rust 1.90, Windows 11 & Ubuntu 24.04, Intel 10700k 8cores
What you did
use std::{
sync::{Arc, atomic::AtomicUsize},
thread,
time::{Duration, Instant},
};
use bevy_app::{
App, Main, TaskPoolOptions, TaskPoolPlugin, TaskPoolThreadAssignmentPolicy, Update,
};
use bevy_ecs::schedule::ExecutorKind;
use std::sync::atomic::Ordering::Relaxed;
const SYSTEM_COUNT: usize = 8;
fn main() {
let mut schedule = App::new();
println!("{}", bevy_tasks::available_parallelism());
let mut options = TaskPoolOptions::with_num_threads(bevy_tasks::available_parallelism());
options.io.min_threads = 0;
options.async_compute.min_threads = 0;
options.io.max_threads = 0;
options.async_compute.max_threads = 0;
options.compute = TaskPoolThreadAssignmentPolicy {
min_threads: 1,
max_threads: bevy_tasks::available_parallelism(),
percent: 1.0,
on_thread_spawn: Some(Arc::new(|| {
println!("spawned ");
})),
on_thread_destroy: None,
};
schedule.add_plugins(TaskPoolPlugin {
task_pool_options: options,
});
let t0 = Instant::now();
let _ = bevy_tasks::ComputeTaskPool::get().scope(|s| {
for _ in 0..SYSTEM_COUNT {
s.spawn(async {
thread::sleep(Duration::from_millis(10));
});
}
});
let duration = t0.elapsed();
println!(" taskpool time consumed: {duration:?}");
for _i in 0..SYSTEM_COUNT {
schedule.add_systems(Update, || {
thread::sleep(Duration::from_millis(10));
});
}
schedule
.get_schedule_mut(Update)
.unwrap()
.set_executor_kind(ExecutorKind::MultiThreaded);
schedule
.get_schedule_mut(Main)
.unwrap()
.set_executor_kind(ExecutorKind::MultiThreaded);
let t0 = Instant::now();
schedule.update();
let duration = t0.elapsed();
println!(" schedule time consumed: {duration:?}");
}What went wrong
- what were you expecting?
The time consumed should be around 10-20ms with parallel execution. - what actually happened?
- The output indicates the taskpool and systems are serial:
16
taskpool time consumed: 80.668829ms
schedule time consumed: 81.100603mswith
println!(
"available threads: {}",
bevy_tasks::ComputeTaskPool::get().thread_num()
);The available thread is 1 even I set so many options. Apparently it is default to a single thread task pool and the whole api is misleading.