#thread #parallel #join #performance

deprecated rayon-futures

(deprecated) Futures integration into Rayon

3 unstable releases

Uses old Rust 2015

0.1.1 Dec 22, 2019
0.1.0 Nov 9, 2017
0.0.1 Sep 9, 2017

#78 in #join

Apache-2.0/MIT

300KB
5K SLoC

Future integration into Rayon

This crate is now deprecated, because it only supports the obsolete futures-0.1. New integration with std::future::Future will likely take place directly in rayon-core, but this is left to the ... future.

NOTE: rayon-futures currently requires unstable features of rayon-core, which may only be enabled with rustc --cfg, e.g. by setting RUSTFLAGS=--cfg rayon_unstable in the environment.

How futures work

Let's start with a brief coverage of how futures work. Our example will be a simple chain of futures:

F_map -> F_socket

Here F_socket is a future that maps to a TCP socket. It returns a Vec<u8> of data read from that socket. F_map is a future will take that data and do some transformation. (Note that the real futures for reading from sockets etc do not work in this way, this is just an example.)

The idea of futures is that each future offers a poll() method. When poll() is invoked, the future will attempt to execute. Typically, this often involves recursively calling poll() on other futures. So, in our example, F_map when it starts would call F_socket.poll() to see if the data is ready. The idea is that poll() returns one of three values:

  • Ok(Async::Ready(R)) -- the future has completed, here is the result R.
  • Err(E) -- the future has completed and resulted in an error E.
  • Ok(Async::NotReady) -- the future is not yet complete.

The last one is the most interesting. It means that the future is blocked on some event X, typically an I/O event (i.e., we are waiting for more data to arrive on a TCP socket).

When a future returns NotReady, it also has one additional job. It must register the "current task" (think for now of the current thread) to be re-awoken when the event X has occurred. For most futures, this job is delegated to another future: e.g., in our example, F_map invokes F_socket.poll(). So if F_socket.poll() returns not-ready, then it will have registered the current thread already, and hence F_map can merely propagates the NotReady result further up.

The current task and executor

A key concept of the futures.rs library is that of an executor. The executor is the runtime that first invokes the top-level future (F_map, in our example). This is precisely the role that Rayon plays. Note that in any futures system there may be many interoperating executors though.

Part of an executor's job is to maintain some thread-local storage (TLS) when a future is executing. In particular, it must setup the "current task" (basically a unique integer, although it's an opaque type) as well as an "unpark object" of type Arc<Unpark>. The Unpark trait offers a single method (unpark()) which can be invoked when the task should be re-awoken. So F_socket might, for example, get the current Arc<Unpark> object and store it for use by an I/O thread. The I/O thread might invoke epoll() or select() or whatever and, when it detects the socket has more data, invoke the unpark() method.

Rayon's futures integration

When you spawn a future of type F into rayon, the idea is that it is going to start independently executing in the thread-pool. Meanwhile, the spawn_future() method returns to you your own future (let's call it F') that you can use to poll and monitor its progress. Internally within Rayon, however, we only allocate a single Arc to represent both of these things -- an Arc<ScopeFuture<F>>, to be precise -- and this Arc hence serves two distinct roles.

The operations on F' (the handle returned to the user) are specified by the trait ScopeFutureTrait and are very simple. The user can either poll() the future, which is checking to see if rayon is done executing it yet, or cancel() the future. cancel() occurs when F' is dropped, which indicates that the user no longer has interest in the result.

Future reference counting

Each spawned future is represented by an Arc. This Arc actually has some interesting structure. Each of the edges in the diagram below represents something that is "kept alive" by holding a ref count (in some way, usually via an Arc):

  F' ---+  [ deque ] --+
        |              |
        v              v
  +---> /---------------------\
  |     | registry:           | ------> [rayon registry]
  |     | contents: --------\ |
  |     | | scope           | | ------> [spawning scope]
  |     | | this            | | --+ (self references)
  |     | | ...             | |   |
  |     | \-----------------/ |   |
  |     \---------------------/   |
  +-------------------------------+

Let's walk through them:

  • The incoming edge from F' represents the edge from the future that was returned to the caller of spawn_future. This ensures that the future arc will not be freed so long as the caller is still interested in looking at its result.
  • The incoming edge from [ deque ] represents the fact that when the future is enqueued into a thread-local deque (which it only sometimes is), that deque holds a ref. This is done by transmuting the Arc into a *const Job object (and hence the *const logically holds the ref that was owned by the Arc). When the job is executed, it is transmuted back and the resulting Arc is eventually dropped, releasing the ref.
  • The registry field holds onto an Arc<Registry> and hence keeps some central registry alive. This doesn't really do much but prevent the Registry from being dropped. In particular, this doesn't prevent the threads in a registry from terminating while the future is unscheduled etc (though other fields in the future do).
  • The scope field (of type S) is the "enclosing scope". This scope is an abstract value that implements the FutureScope<'scope> trait -- this means that it is responsible for ensuring that 'scope does not end until one of the FutureScope methods are invoked (which occurs when the future has finished executing). For example, if the future is spawned inside a scope() call, then the S will be a wrapper (ScopeFutureScope) around a *const Scope<'scope>. When the future is created one job is allocated for this future in the scope, and the scope counter is decremented once the future is marked as completing.
    • In general, the job of the scope field is to ensure that the future type (F) remains valid. After all, since F: 'scope, F is known to be valid until the lifetime 'scope ends, and that lifetime cannot end until the scope methods are invoked, so we know that F must stay valid until one of those methods are invoked.
    • All of our data of type F is stored in the field spawn (not shown here). This field is always set to None before the scope counter is decremented. See the section on lifetime safety for more details.
  • The this field stores an Arc which is actually this same future. Thus the future has a ref count cycle and cannot be freed until this cycle is broken. That field is actually an Option<Arc<..>> and will be set to None once the future is complete, breaking the cycle and allowing it to be freed when other references are dropped.

The future state machine

Internally, futures go through various states, depicted here:

PARKED <----+
|           |
v           |
UNPARKED    |
|           |
v           |
EXECUTING --+
|   |   ^
|   v   |
|   EXECUTING_UNPARKED
|
v
COMPLETE

When they are first created, futures begin as PARKED. A PARKED future is one that is waiting for something to happen. It is not scheduled in the deque of any thread. Even before we return from spawn_future(), however, we will transition into UNPARKED. An UNPARKED future is one that is waiting to be executed. It is enqueued in the deque of some Rayon thread and hence will execute when the thread gets around to it.

Once the future begins to execute (it itself is a Rayon job), it transitions into the EXECUTING state. This means that it is busy calling F.poll(), basically. While it calls poll(), it also sets up its contents.this field as the current "notify" instance. Hence if F returns NotReady, it will clone the this field and hold onto it to signal us the future is ready to execute again.

For now let's assume that F is complete and hence returns either Ok(Ready(_)) or Err(_). In that case, the future can transition to COMPLETE. At this point, many bits of state that are no longer needed (e.g., the future itself, but also the this field) are set to None and dropped, and the result is stored in the result field. (Moreover, we may have to signal other tasks, but that is discussed in a future section.)

If F returns Ok(Async::NotReady), then we would typically transition to the PARKED state and await the call to notify(). When notify() is called, it would move the future into the UNPARK state and inject it into the registry.

However, due to the vagaries of thread-scheduling, it can happen that notify() is called before we exit the EXECUTING state. For example, we might invoke F.poll(), which sends the Unpark instance to the I/O thread, which detects I/O, and invokes notify(), all before F.poll() has returned. In that case, the notify() method will transition the state (atomically, of course) to EXECUTING_UNPARKED. In that case, instead of transitioning to PARKED when F.poll() returns, the future will simply transition right back to EXECUTING and try calling poll() again. This can repeat a few times.

Lifetime safety

Of course, Rayon's signature feature is that it allows you to use a future F that includes references, so long as those references outlive the lifetime of the scope 'scope. So why is this safe?

The basic idea of why this is safe is as follows. The ScopeFuture struct holds a ref on the scope itself (via the field scope). Until this ref is decremented, the scope will not end (and hence 'scope is still active). This ref is only decremented while the future transitions into the COMPLETE state -- so anytime before then, we know we don't have to worry, the references are still valid.

As we transition into the COMPLETE state is where things get more interesting. You'll notice that signaling the self.scope job as done is the last thing that happens during that transition. Importantly, before that is done, we drop all access that we have to the type F: that is, we store None into the fields that might reference values of type F. This implies that we know that, whatever happens after we transition into COMPLETE, we can't access any of the references found in F anymore.

This is good, because there are still active refs to the ScopeFuture after we enter the COMPLETE state. There are two sources of these: unpark values and the future result.

NotifyHandle values. We may have given away NotifyHandle values -- these contain trait objects that are actually refs to our ScopeFuture. Note that NotifyHandle: 'static, so these could be floating about for any length of time (we had to transmute away the lifetimes to give them out). This is ok because (a) the Arc keeps the ScopeFuture alive and (b) the only thing you can do is to call notify(), which will promptly return since the state is COMPLETE (and, anyhow, as we saw above, it doesn't have access to any references anyhow).

Future result. The other, more interesting reference to the ScopeFuture is the value that we gave back to the user when we spawned the future in the first place. This value is more interesting because it can be used to do non-trivial things, unlike the NotifyHandle. If you look carefully at this handle, you will see that its type has been designed to hide the type F. In fact, it only reveals the types T and E which are the ok/err result types of the future F. This is intentonal: suppose that the type F includes some references, but those references don't appear in the result. We want the "result" future to be able to escape the scope, then, to any place where the types T and E are still in scope. If we exposed F here that would not be possible. (Hiding F also requires a transmute to an object type, in this case an internal trait called ScopeFutureTrait.) Note though that it is possible for T and E to have references in them. They could even be references tied to the scope.

So what can a user do with this result future? They have two operations available: poll and cancel. Let's look at cancel first, since it's simpler. If the state is COMPLETE, then cancel() is an immediate no-op, so we know that it can't be used to access any references that may be invalid. In any case, the only thing it does is to set a field to true and invoke notify(), and we already examined the possible effects of notify() in the previous section.

So what about poll()? This is how the user gets the final result out of the future. The important thing that it does is to access (and effectively nullify) the field result, which stores the result of the future and hence may have access to T and E values. These values may contain references...so how do we know that they are still in scope? The answer is that those types are exposed in the user's type of the future, and hence the basic Rust type system should guarantee that any references are still valid, or else the user shouldn't be able to call poll(). (The same is true at the time of cancellation, but that's not important, since cancel() doesn't do anything of interest.)

Dependencies

~650KB