Struct rayon::ThreadPoolBuilder

pub struct ThreadPoolBuilder<S = DefaultSpawn> {
    pub(crate) num_threads: usize,
    pub(crate) use_current_thread: bool,
    pub(crate) panic_handler: Option<Box<dyn Fn(Box<dyn Any + Send>) + Sync + Send>>,
    pub(crate) get_thread_name: Option<Box<dyn FnMut(usize) -> String>>,
    pub(crate) stack_size: Option<usize>,
    pub(crate) start_handler: Option<Box<dyn Fn(usize) + Sync + Send>>,
    pub(crate) exit_handler: Option<Box<dyn Fn(usize) + Sync + Send>>,
    pub(crate) spawn_handler: S,
    pub(crate) breadth_first: bool,
}

Used to create a new ThreadPool or to configure the global rayon thread pool.

Creating a ThreadPool

The following creates a thread pool with 22 threads.

let pool = rayon::ThreadPoolBuilder::new().num_threads(22).build().unwrap();

To instead configure the global thread pool, use build_global():

rayon::ThreadPoolBuilder::new().num_threads(22).build_global().unwrap();

Fields

num_threads: usize

use_current_thread: bool

panic_handler: Option<Box<dyn Fn(Box<dyn Any + Send>) + Sync + Send>>

get_thread_name: Option<Box<dyn FnMut(usize) -> String>>

stack_size: Option<usize>

start_handler: Option<Box<dyn Fn(usize) + Sync + Send>>

exit_handler: Option<Box<dyn Fn(usize) + Sync + Send>>

spawn_handler: S

breadth_first: bool

Implementations

impl ThreadPoolBuilder

pub fn new() -> ThreadPoolBuilder

Creates and returns a valid rayon thread pool builder, but does not initialize it.

impl<S> ThreadPoolBuilder<S>

where S: ThreadSpawn,

Note: the S: ThreadSpawn constraint is an internal implementation detail for the default spawn and those set by spawn_handler.

pub fn build(self) -> Result<ThreadPool, ThreadPoolBuildError>

Creates a new ThreadPool initialized using this configuration.

pub fn build_global(self) -> Result<(), ThreadPoolBuildError>

Initializes the global thread pool. This initialization is optional. If you do not call this function, the thread pool will be automatically initialized with the default configuration. Calling build_global is not recommended, except in two scenarios:

• You wish to change the default configuration.
• You are running a benchmark, in which case initializing may yield slightly more consistent results, since the worker threads will already be ready to go even in the first iteration. But this cost is minimal.

Initialization of the global thread pool happens exactly once. Once started, the configuration cannot be changed. Therefore, if you call build_global a second time, it will return an error. An Ok result indicates that this is the first initialization of the thread pool.

impl ThreadPoolBuilder

pub fn build_scoped<W, F, R>( self, wrapper: W, with_pool: F, ) -> Result<R, ThreadPoolBuildError>

where W: Fn(ThreadBuilder) + Sync, F: FnOnce(&ThreadPool) -> R,

Creates a scoped ThreadPool initialized using this configuration.

This is a convenience function for building a pool using std::thread::scope to spawn threads in a spawn_handler. The threads in this pool will start by calling wrapper, which should do initialization and continue by calling ThreadBuilder::run().

Examples

A scoped pool may be useful in combination with scoped thread-local variables.

scoped_tls::scoped_thread_local!(static POOL_DATA: Vec<i32>);

fn main() -> Result<(), rayon::ThreadPoolBuildError> {
    let pool_data = vec![1, 2, 3];

    // We haven't assigned any TLS data yet.
    assert!(!POOL_DATA.is_set());

    rayon::ThreadPoolBuilder::new()
        .build_scoped(
            // Borrow `pool_data` in TLS for each thread.
            |thread| POOL_DATA.set(&pool_data, || thread.run()),
            // Do some work that needs the TLS data.
            |pool| pool.install(|| assert!(POOL_DATA.is_set())),
        )?;

    // Once we've returned, `pool_data` is no longer borrowed.
    drop(pool_data);
    Ok(())
}

impl<S> ThreadPoolBuilder<S>

pub fn spawn_handler<F>(self, spawn: F) -> ThreadPoolBuilder<CustomSpawn<F>>

where F: FnMut(ThreadBuilder) -> Result<(), Error>,

Sets a custom function for spawning threads.

Note that the threads will not exit until after the pool is dropped. It is up to the caller to wait for thread termination if that is important for any invariants. For instance, threads created in std::thread::scope will be joined before that scope returns, and this will block indefinitely if the pool is leaked. Furthermore, the global thread pool doesn’t terminate until the entire process exits!

Examples

A minimal spawn handler just needs to call run() from an independent thread.

fn main() -> Result<(), rayon::ThreadPoolBuildError> {
    let pool = rayon::ThreadPoolBuilder::new()
        .spawn_handler(|thread| {
            std::thread::spawn(|| thread.run());
            Ok(())
        })
        .build()?;

    pool.install(|| println!("Hello from my custom thread!"));
    Ok(())
}

The default spawn handler sets the name and stack size if given, and propagates any errors from the thread builder.

fn main() -> Result<(), rayon::ThreadPoolBuildError> {
    let pool = rayon::ThreadPoolBuilder::new()
        .spawn_handler(|thread| {
            let mut b = std::thread::Builder::new();
            if let Some(name) = thread.name() {
                b = b.name(name.to_owned());
            }
            if let Some(stack_size) = thread.stack_size() {
                b = b.stack_size(stack_size);
            }
            b.spawn(|| thread.run())?;
            Ok(())
        })
        .build()?;

    pool.install(|| println!("Hello from my fully custom thread!"));
    Ok(())
}

This can also be used for a pool of scoped threads like crossbeam::scope, or std::thread::scope introduced in Rust 1.63, which is encapsulated in build_scoped.

fn main() -> Result<(), rayon::ThreadPoolBuildError> {
    std::thread::scope(|scope| {
        let pool = rayon::ThreadPoolBuilder::new()
            .spawn_handler(|thread| {
                let mut builder = std::thread::Builder::new();
                if let Some(name) = thread.name() {
                    builder = builder.name(name.to_string());
                }
                if let Some(size) = thread.stack_size() {
                    builder = builder.stack_size(size);
                }
                builder.spawn_scoped(scope, || {
                    // Add any scoped initialization here, then run!
                    thread.run()
                })?;
                Ok(())
            })
            .build()?;

        pool.install(|| println!("Hello from my custom scoped thread!"));
        Ok(())
    })
}

pub fn thread_name<F>(self, closure: F) -> ThreadPoolBuilder<S>

where F: FnMut(usize) -> String + 'static,

Sets a closure which takes a thread index and returns the thread’s name.

pub fn num_threads(self, num_threads: usize) -> ThreadPoolBuilder<S>

Sets the number of threads to be used in the rayon threadpool.

If you specify a non-zero number of threads using this function, then the resulting thread-pools are guaranteed to start at most this number of threads.

If num_threads is 0, or you do not call this function, then the Rayon runtime will select the number of threads automatically. At present, this is based on the RAYON_NUM_THREADS environment variable (if set), or the number of logical CPUs (otherwise). In the future, however, the default behavior may change to dynamically add or remove threads as needed.

Future compatibility warning: Given the default behavior may change in the future, if you wish to rely on a fixed number of threads, you should use this function to specify that number. To reproduce the current default behavior, you may wish to use std::thread::available_parallelism to query the number of CPUs dynamically.

Old environment variable: RAYON_NUM_THREADS is a one-to-one replacement of the now deprecated RAYON_RS_NUM_CPUS environment variable. If both variables are specified, RAYON_NUM_THREADS will be preferred.

pub fn use_current_thread(self) -> ThreadPoolBuilder<S>

Use the current thread as one of the threads in the pool.

The current thread is guaranteed to be at index 0, and since the thread is not managed by rayon, the spawn and exit handlers do not run for that thread.

Note that the current thread won’t run the main work-stealing loop, so jobs spawned into the thread-pool will generally not be picked up automatically by this thread unless you yield to rayon in some way, like via yield_now(), yield_local(), or scope().

Local thread-pools

Using this in a local thread-pool means the registry will be leaked. In future versions there might be a way of cleaning up the current-thread state.

pub fn panic_handler<H>(self, panic_handler: H) -> ThreadPoolBuilder<S>

where H: Fn(Box<dyn Any + Send>) + Send + Sync + 'static,

Normally, whenever Rayon catches a panic, it tries to propagate it to someplace sensible, to try and reflect the semantics of sequential execution. But in some cases, particularly with the spawn() APIs, there is no obvious place where we should propagate the panic to. In that case, this panic handler is invoked.

If no panic handler is set, the default is to abort the process, under the principle that panics should not go unobserved.

If the panic handler itself panics, this will abort the process. To prevent this, wrap the body of your panic handler in a call to std::panic::catch_unwind().

pub fn stack_size(self, stack_size: usize) -> ThreadPoolBuilder<S>

Sets the stack size of the worker threads

pub fn breadth_first(self) -> ThreadPoolBuilder<S>

👎Deprecated: use scope_fifo and spawn_fifo for similar effect

(DEPRECATED) Suggest to worker threads that they execute spawned jobs in a “breadth-first” fashion.

Typically, when a worker thread is idle or blocked, it will attempt to execute the job from the top of its local deque of work (i.e., the job most recently spawned). If this flag is set to true, however, workers will prefer to execute in a breadth-first fashion – that is, they will search for jobs at the bottom of their local deque. (At present, workers always steal from the bottom of other workers’ deques, regardless of the setting of this flag.)

If you think of the tasks as a tree, where a parent task spawns its children in the tree, then this flag loosely corresponds to doing a breadth-first traversal of the tree, whereas the default would be to do a depth-first traversal.

Note that this is an “execution hint”. Rayon’s task execution is highly dynamic and the precise order in which independent tasks are executed is not intended to be guaranteed.

This breadth_first() method is now deprecated per RFC #1, and in the future its effect may be removed. Consider using scope_fifo() for a similar effect.

pub fn start_handler<H>(self, start_handler: H) -> ThreadPoolBuilder<S>

where H: Fn(usize) + Send + Sync + 'static,

Sets a callback to be invoked on thread start.

The closure is passed the index of the thread on which it is invoked. Note that this same closure may be invoked multiple times in parallel. If this closure panics, the panic will be passed to the panic handler. If that handler returns, then startup will continue normally.

pub fn exit_handler<H>(self, exit_handler: H) -> ThreadPoolBuilder<S>

where H: Fn(usize) + Send + Sync + 'static,

Sets a callback to be invoked on thread exit.

The closure is passed the index of the thread on which it is invoked. Note that this same closure may be invoked multiple times in parallel. If this closure panics, the panic will be passed to the panic handler. If that handler returns, then the thread will exit normally.

Trait Implementations

impl<S> Debug for ThreadPoolBuilder<S>

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Default for ThreadPoolBuilder

fn default() -> ThreadPoolBuilder

Returns the “default value” for a type.