Tokio asynchronous I/O examples

While studying the inner workings of Tokio, Mio, and futures, I developed several small example programs to better understand usage and explore somewhat different approaches or scenarios than the example programs provided with tokio-core. These example programs are based on (at most) tokio-core and not higher-level crates like tokio-proto and tokio-service.

I documented my conclusions from this study in a blog post:

Tokio internals: Understanding Rust's asynchronous I/O framework from the bottom up

UDP echo examples

The tokio-core echo-udp.rs example program operates in a flip-flop fashion where it is either listening for an incoming datagram, or sending an outgoing datagram. While it is waiting for a send of an outgoing datagram to complete, it cannot process any pending incoming datagrams. This flip-flop behavior is probably fine for many protocols like DNS and NTP, since a certain amount of incoming packets will be buffered in the kernel queue. Other more complex protocols or scenarios may require true multiplexing of reads and writes.

I wrote a series of small UDP echo example programs that listen on IPv4 localhost port 2000, maintain a small outgoing queue for the echos, and multiplex reads and writes. I started with implementations based on the operating system select() and epoll facilities, and worked my way up to Mio and Tokio implementations.

• echo-select.rs: This implementation uses the select() system call to manage I/O. This only works on systems supporting select() (Linux, Mac OS, etc.) and is only compiled when the select feature flag is given.

• echo-epoll-level.rs: This implementation uses the Linux epoll facility in level-triggered mode as a "better select".

• echo-epoll-edge.rs: This implementation uses the Linux epoll facility in edge-triggered mode. Edge-triggered events can theoretically provide better performance than level-triggering by reducing the overhead related to selection. Handlers are expected to perform as much I/O as possible until an EWOULDBLOCK is indicated, at which time epoll_wait() is called again and other file descriptors may be handled. Any mitigation of the edge-triggered starvation problem is up to the application, and no such mitigation is demonstrated here.

• echo-mio-level.rs: A simple UDP echo server using the cross-platform mio crate to multiplex reads and writes. This program uses level-triggered events.

• echo-mio-edge.rs: A simple UDP echo server using the mio crate to multiplex reads and writes. This program uses edge-triggered events, which can theoretically provide better performance than level-triggering by reducing the overhead related to selection. Handlers are expected to perform as much I/O as possible until WouldBlock is indicated, at which time Poll::poll() is called again and other file descriptors may be handled. Any mitigation of the edge-triggered starvation problem is up to the application, and no such mitigation is demonstrated here.

• echo-tokio.rs: A simple UDP echo server using Tokio to multiplex reads and writes.

• echo-tokio-mpsc.rs: A simple UDP echo server using Tokio to multiplex reads and writes. This is an alternate implementation that uses separate "reader" and "writer" futures connected by an MPSC queue.

Futures and task notification

• future-notify.rs: This small program demonstrates how futures can be manually scheduled for polling by calling the notify() method on their task.

Understanding Mio

• mio-empty.rs: This is an "empty" Mio example. Mio is polled without having registered for any events, so the poll() never returns. This can be useful for studying basic Mio behaviors that occur regardless of any registrations. For example, when run via strace on Linux, we can see that Mio always creates a pipe to accommodate non-system events sourced from user-space, and then configures the underlying epoll to watch for read events on the pipe.

• mio-mixed.rs: This program demonstrates how a single Mio instance can be used to receive both system events (e.g. file descriptor events) and non-system events (e.g. events sourced on user-space threads other than the thread running the Mio poll). We listen for incoming UDP datagrams on port 2000, and also listen for events created by our timer thread every three seconds. Running this program on Linux via strace shows how Mio notifies the polling thread of the non-system event by writing to a pipe.

• mio-pipe.rs: Demonstrate a possible bug where Mio uses a pipe write to notify of a mio::Registration event which occurs while epoll_wait() is not happening. For more details, see: tokio-rs/mio#785

Multiple sockets in Tokio

There are several ways to have Tokio manage multiple sockets.

These programs listen for incoming UDP datagrams on IPv4 localhost ports 2000 through 2009, and print a summary of each datagram to the standard output.

• tokio-multisocket-join.rs: This implementation works by creating ten futures, each processing data on one socket, and combining them into a single composite future via join_all(). This composite future is provided to Tokio via Core::run(). This is a simple approach, but a possible downside is that every future is polled whenever a single packet arrives on a socket. This is because all the futures run within a single task. Because notifications happen at the task level, any notification arranged in any of the futures will cause the main task to be notified. It will poll the top-level FromAll future, which itself will poll each of its children.

• tokio-multisocket-spawn.rs: This implementation works by creating ten futures, each added to the event loop within a distinct task via Handle::spawn(). This spawning is performed by a UdpMultiServer future which is passed to Tokio as the main future via Core::run(). In contrast to tokio-multisocket-join.rs, this approach avoids polling all futures when only a single future needs to be polled.

• tokio-multisocket-futuresunordered.rs: This implementation works by creating ten futures, each processing data on one socket, and managing them with a FuturesUnordered stream, which is provided to Tokio's Core::run() by way of the stream's for_each() method. FuturesUnordered has a very useful property that makes it potentially more attractive (in this scenario) than a simple join. A Join future, when polled, will in turn poll all of its active futures, even if only one future needs to be polled (i.e., only one future arranged a notification event for the task which contains the Join and its child futures). In contrast, "Futures managed by FuturesUnordered will only be polled when they generate notifications". This is accomplished by FuturesUnordered::poll() polling each future with a distinct NotifyHandle thread-local, so it can perform per-future notification discrimination and only poll the futures that need to be polled. When this program is run, you can observe that only the correct future is polled.

Building

The example programs can be built with cargo build. The select() and epoll examples may be built on suitable platforms if the select and/or epoll feature flags are enabled. For example:

cargo build --features=select,epoll

License

This project is licensed under either of

• Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option. (The same license terms as Tokio itself.)