Posts Tagged ‘communication’

A lot of people have been asking me how to use protocols in Rust lately, so I thought I’d write up a little tutorial. Custom protocols are how to get the biggest benefits from Rust’s communication system, as this is how you get the biggest safety guarantees and expose the most opportunities for optimization. It’s also more labor intensive, so some of the other library features such as streams or port sets might be a better starting point. This post, however, introduces how to write your own protocols.

Protocols are created using the proto! syntax extension. Each protocol has a collection of states, and each state in a protocol has a collection of messages that are allowed. Every protocol must have at least states, though states needn’t have any messages. Thus, we will start with the simplest possible protocol definition:

proto! simple_proto {
    StartState:send { }
}

This creates a protocol called simple_proto, with a single state called StartState. Protocols always start in the first state listed. There is one other decoration, :send. This indicates that StartState is a send state. Protocols are always written from the client perspective, so a send state means it is a state in which the client may send a message. The protocol compiler will generate a dual state for the server, meaning there will be a version of StartState that is expecting to receive a message.

Given this specification, the protocol compiler will create a module called simple_proto, that includes types and functions for communicating according to this protocol. One generated function is called init, which is used to create a pipe in the start state. We can start the protocol like this:

let (server, client) = simple_proto::init();

This creates a pair of endpoints. The client endpoint will have the type simple_proto::client::StartStart, meaning it is in the client’s version of StartState. The server endpoint, on the other hand, has type simple_proto::server::StartState, which means it is in the server’s version of the StartState. The difference means that client is expecting to send a message, while server is expecting to receive a message.

We can’t really do anything further with this protocol, since there are no messages defined. Strictly speaking, the server could try to receive, but it would block forever since there is no way for the sender to send a message. Let’s fix this by adding a message.

proto! simple_proto {
    StartState:send {
        SayHello -> StartState
    }
}

Messages have the form Name(arguments) -> NextState. In this case, we added a message called SayHello, which carries no data with it. After sending a SayHello message, the protocol transitions to (or stays in, really) the StartState. We can now write some functions that communicate with each other. Here’s an example client program:

fn client(+channel: simple_proto::client::StartState) {
    import simple_proto::client;

    client::SayHello(channel);
}

Receive, by itself, is a little trickier. I recommend using the select macro instead. For now, you can get the select macro here. Once macro import is working, the select macro will be included in the standard library. For now, to use the select macro, save it in a file called select-macro.rs and add the following lines near the top of your program.

fn macros() {
    include!("select-macro.rs");
}

Once you’ve done this, you can write the server as follows.

fn server(+channel: simple_proto::server::StartState) {
    import simple_proto::SayHello;

    select! {
        channel => {
            SayHello -> _channel {
                io::println("Client says hello!");
            }
        }
    }
}

Select allows you to provide a set of endpoints to listen for messages on, followed by actions to take depending on which one receives a message. In this case, we only have one endpoint, called channel, which is in the server StartState state. After the =>, there is a block describing message patterns and code to execute if the pattern is matched. In this case, we only have one pattern, SayHello -> _channel. This mirrors the definition of the message in the protocol specification. It says “if we receive a SayHello message, bind an endpoint in the next protocol state to _channel and execute the code in the following block.” We use _channel for the next state because in this case we are not planning on sending or receiving any more messages.

Now let’s make this protocol a little more interesting by adding a new state and a message that carries data. We will do this by letting the client ask the server’s name and wait for a reply. The new protocol looks like this:

proto! simple_proto {
    StartState:send {
        SayHello -> StartState,
        WhatsYourName -> GettingName
    }

    GettingName:recv {
        MyNameIs(~str) -> StartState
    }
}

We’ve added a new message to StartState, which the client uses to ask the server’s name. After sending this, the protocol transitions to the GettingName state, where the client will wait to receive the MyNameIs message from the server. At this point, the protocol moves back to the StartState, and we can do it all over again. We’ve added an argument to MyNameIs, which means this message carries a string with it. Our client code now looks like this:

fn client(+channel: simple_proto::client::StartState) {
    import simple_proto::client;
    import simple_proto::MyNameIs;

    let channel = client::SayHello(channel);
    let channel = client::WhatsYourName(channel);
    select! {
        channel => {
            MyNameIs(name) -> _channel {
                io::println(fmt!("The server is named %s", *name));
            }
        }
    }
}

At a high level, this code says hello to the server, then asks for it’s name, then waits for the response and reports the server’s name to the user. It probably looks a little add that every line starts with let channel = .... This is because endpoints are single use. Any time you send or receive a message on an endpoint, the endpoint is consumed. Fortunately, all the send and receive functions return a new endpoint that you can use to continue the protocol.

The use of select! here is similar to how it was in the previous server example, except that we’ve added name to the MyNameIs pattern. This matches the ~str parameter in the protocol specification, and it binds the string sent by the server to name, so that we can print it out in the handler code.

For the new server, we need to add another clause to the message patterns:

fn server(+channel: simple_proto::server::StartState) {
    import simple_proto::{SayHello, WhatsYourName};
    import simple_proto::server::MyNameIs;

    select! {
        channel => {
            SayHello -> _channel {
                io::println("Client says hello!");
            },
            WhatsYourName -> channel {
                MyNameIs(channel, ~"Bob");
            }
        }
    }
}

In this case, if we receive a WhatsYourName message, we send a MyNameIs message on the new endpoint (called channel), which contains the string ~"Bob", which is what this server has decided to call itself. The client will eventually receive this string and show it to the user.

This covers the basic definition and usage of protocols. There are several other features, however. This includes polymorphic states and terminal states. Polymorphic states allow you to create protocols that work for different types. One common example is the stream protocol, which lets you send a whole bunch of messages of a given type:

proto! stream {
    Stream:send<T:send> {
        Send(T) -> Stream<T>
    }
}

We can add as many type parameters as we want to each of the states, with arbitrary bounds as well. You’ll probably want all your data types to be send-bounded though. Then, each time a message transitions to a polymorphic state, it must provide type parameters. You can see this on the right had side of the Send message.

Sometimes, we want to have a message that ends the protocol. For example, in our previous example, we might want a GoodBye message. One way to do this is to make a state with no messages, and step to that:

proto! simple_proto {
    StartState:send {
        GoodBye -> Done,
    }

    Done { }
}

However, this is a little verbose, and it also hides that fact that you really intended the protocol to end there. Thus, there is a special form that indicates sending a message ends the protocol. We write it like this:

proto! simple_proto {
    StartState:send {
        GoodBye -> !
    }
}

Stepping to ! represents closing the protocol and is analogous to how a function that returns ! actually never returns. When a message steps to !, neither the corresponding send function nor the receive function will return a new endpoint, meaning there is no way you could even attempt to send messages on this connection.

I hope this helps to get started with protocols in Rust. There are a few other features, but this covers the basics. Please feel free to ask questions!

Advertisements

The benchmarks in my last post had one thing in common: all communication was one sender to one receiver. It’s surprising how often this is sufficient, but sooner or later we are going to need a way to have multiple tasks sending to the same receiver. I’ve been experimenting with two ways of doing many senders to different receivers, and I now have some results to show.

The pipes library includes a select operation. This lets you listen on several receive endpoints simultaneously. Unfortunately, the single-use nature of endpoints makes select a little clunky to use. To help alleviate this, I added a port_set to the library. Port sets allow you to easily treat several receive endpoints as a unit. This allows send to still be very fast, but receive is a little bit slower due to the overhead setting up and tearing down the select operation. The current implementation for select is O(n) in the number of endpoints, so this works well for small numbers of tasks, but breaks down as things get bigger.

The other option is to slow down the sending end, using something I call a shared_chan. This is a send endpoint wrapped in an exclusive ARC. Now all the senders have to contend with each other, but the receive side is exactly as cheap as before. For cases where you have a lot of senders that send messages relatively infrequently, this will likely outperform the port_set approach, at least until select is faster.

Both of these are sufficient to run the msgsend benchmark that I talked about at the beginning of all of this. Here are the results, combined with the previous numbers.

Language Messages per second Comparison
Rust port_set 881,578 232.8%
Scala 378,740 100.0%
Rust port/chan (updated) 227,020 59.9%
Rust shared_chan 173,436 45.8%
Erlang (Bare) 78,670 20.8%
Erlang (OTP) 76,405 20.2%

The most obvious thing is that the port_set version is over twice as fast as Scala, the previous winner. I also re-ran the port/chan version for comparison, and it got a little bit faster. There has been quite a bit of churn in Rust recently, so it’s quite possible that these showed up here as better performance.

Writing the port_set version proved the most interesting to me. Relying on select ended up relaxing some of the ordering guarantees. Previously if we had Task A send a message to Task C and then send a message to Task B, and then have Task B wait to receive message to from Task A and then send a message to Task C, we could count on Task C seeing Task A’s message before seeing Task B’s message. With the port_set, this is no longer true, although we still preserve the order in messages sent by a single task. An easy way to work around this, however, was to rely on pipe’s closure reporting ability. The server could tell when a worker would no longer send any more messages because it would detect when the worker closed its end of the pipe.

I hinted in my last post that pipes in Rust have very good performance. This falls out of the fact that the protocol specifications provide very strong static guarantees about what sorts of things can happen at runtime. This allows, among other things, for message send/receive fastpath that requires only two atomic swaps.

Let’s start with the message ring benchmark. I posted results from this earlier. This benchmark spins up a bunch of tasks that arrange themselves in a while. Each task sends a message to their right-hand neighbor, and receives a message from the left-hand neighbor. This repeats for a while. At the end, we look at the total time taken divided by the number of messages. This gives us roughly the fastest we can send and receive a message, modulo some task spawning overhead. The existing port/chan system was able to send about 250,000 messages per second, or one message every 3.9 µs. Here are the results for pipes:

Sent 1000000 messages in 0.227634 seconds
  4.39301e+06 messages / second
  0.227634 µs / message

This is about 17x faster!

It would be a bit dishonest to stop here, however. I wrote this benchmark specifically to make any new implementation really shine. The question is whether faster message passing makes a difference on bigger programs.

To test this, I started by updating the Graph500 Parallel Breadth First Search benchmark. This code gets its parallelism from std::par::map, which in turn is built on core::future. Future has a very simple parallel protocol; it just spawns a task to compute something, which then sends a single message back to the spawner. Porting this was a relatively small change, yet it got measurable speedups. Here are the results.

Benchmark Port/chan time (s) Pipe time (s) Improvement (%)
Graph500 PBFS 0.914772 0.777784 17.6%

The Rust benchmark suite also includes several benchmarks from the Computer Language Benchmarks Game (i.e. the Programming Language Shootout). Some of these, such as k-nucleotide, use Rust’s parallelism features. I went ahead and ported this benchmark over to use pipes, and there are the results.

Benchmark Port/chan time (s) Pipe time (s) Improvement (%)
Shootout K-Nucleotide 4.335 3.125 38.7%

Not too shabby. I’ve been working on porting other benchmarks as well. Some are more difficult because they do not fit the 1:1 nature of pipes very well. In the case of the shootout-threadring benchmark, it actually got significantly slower when I moved to pipes. The thread ring benchmark seems to mostly be measuring the time to switch between tasks, as only one should be runnable at any given time. My hypothesis is that because message passing got faster, this test now hammers the scheduler synchronization code harder, leading to more slowdown due to contention. We’ll need more testing to know for sure. At any rate, scheduler improvements (such as work stealing, which Ben Blum will be working on) should improve this benchmark as well.

Other than that, I’ve been working on rewriting more Rust code to see how it works with pipes versus ports and chans. It has been particularly informative to try to transition parts of Servo over to using pipes.

About a month ago, I posted that I was going to be working on improving Rust’s message passing performance. I quickly threw together a prototype of a new communication system based on a shared queue protected by a mutex. This was about twice as fast as the existing system, because it removed the global mutex from the messaging paths. This prototype hurt expressiveness somewhat, and still it seemed we could do a lot better.

Rust has some extremely powerful features in its type system. The fact that it can deal with concepts like uniqueness, initialization status, copyability, and other traits mean we can encode some very powerful invariants. Thus, I took some inspiration from the Singularity OS and set out to see if I could encode something like channel contracts in Rust. The result is a proposal for a feature I’m calling pipes.

The way pipes work is that when you create a pipe you get two endpoints that are forever entangled together. One endpoint can send one message, and the other endpoint can receive that one message. Sending and receiving destroys the endpoint, but the operation also produces a new endpoint to continue the communication. Endpoints have a state associated with them, which specifies which messages can be sent or received. This information is encoding in the type system, so Rust can statically guarantee that no task will send a message that is not legal in the given state. Pipes are not copyable; they are always for 1:1 communication. However, endpoints can be sent between tasks.

Critical to pipes are the associated protocol specification. Protocols have two views: the client and the server. Protocols are always written from the perspective of the client. This decision was arbitrary, but in general it makes sense to only write down one side of the protocol. The other perspective is generated by reversing the direction of all the messages. Here’s an example of what I’m envisioning for a protocol specification.

proto! bank {
    login:send {
        login(username, password) -> login_response
    }

    login_response:recv {
        ok -> connected,
        invalid -> login
    }

    connected:send {
        deposit(money) -> connected,
        withdrawal(amount) -> withdrawal_response
    }

    withdrawal_response:recv {
        money(money) -> connected,
        insufficient_funds -> connected
    }
}

This describes the protocol you might use in an online banking situation. The protocol has four states (login, login_response, connected and withdrawal_response), each one annotated with whether the sender is allowed to send or receive in that state. In this case, a client would start out in the login state, where the client can attempt to login with a username and password. After sending a login message, the protocol enters the login_response state, where the server informs the client that either the login succeeded (in which case the protocol transitions to the connected state), or the login failed, in which case the protocol returns to the login state and the client can retry.

From the connected state, the client can try to deposit or withdrawal money. We assume that depositing money never fails, so sending a deposit message results in the protocol staying in the connected state. On the other hand, withdrawal can fail, for example, if the account does not have enough money. To model this, sending a withdrawal message results in the protocol going to the withdrawal_response state. Here, the client waits to either receive the requested money, or for a message saying there was not enough money in the account. In both cases, we end up back in the connected state.

Below is a code example showing how a client might use this protocol.

fn bank_client(+bank: bank::client::login) {
    import bank::*;

    let bank = client::login(bank, "theincredibleholk", "1234");
    let bank = alt recv(bank) {
      some(ok(connected)) {
        #move(connected)
      }
      some(invalid(_)) { fail "login unsuccessful" }
      none { fail "bank closed the connection" }
    };

    let bank = client::deposit(bank, 100.00);
    let bank = client::withdrawal(bank, 50.00);
    alt recv(bank) {
      some(money(m, _)) {
        io::println("Yay! I got money!");
      }
      some(insufficient_funds(_)) {
        fail "someone stole my money"
      }
      none {
        fail "bank closed the connection"
      }
    }
}

All of this code in this posts works on the latest Rust compiler as of this morning. I’ve also started transitioning some of our benchmarks to the new pipe system, and the results have been impressive. I’ll have a post diving into the performance of pipes soon.

Today we’re going to take a brief look at message passing performance in Rust. Rust is already pretty quick on this front, but there are some really obvious possibilities for making it even faster. It looks like I’ll be spending most of the summer working on this, so I thought it’d be good to start out with some baseline performance numbers. Rust’s testsuite has a benchmark called msgsend.rs, which is a microbenchmark for message passing performance. It is based on a benchmark from Paul Keeble that original compared Erlang and Scala. Here’s a table summarizing the results:

Language Messages per second Comparison
Scala 378,740 100.0%
Rust 175,373 46.3%
Erlang (Bare) 78,670 20.8%
Erlang (OTP) 76,405 20.2%

These numbers were generated on an Early 2011 MacBook Pro 2.3 GHz Intel Core i7 with 8GB RAM and Lion 10.7.4. Don’t read too much into these numbers; they are incredibly unscientific. The only real use for them will be to see if running the same benchmark for Rust in a few weeks yields a larger number. It’s also worth noting that my results disagree with the original author, who saw Erlang performing about 6 times faster than Scala. I suspect that Erlang’s message passing may not be as efficient on Mac OS X, but I do not know what system the tests were originally run on.

So, where do we go from here? I mentioned that there are some obvious first steps. The most obvious is that right now sending a message involves taking a global lock. In effect, this means we can only send one message in a time. Suppose we had four tasks, A, B, C and D, where A was sending to B and C to D. Intuitively, these two sends should be completely independent of each other, but this is not currently the case in Rust. This is one thing I hope to change.

It would be nice if we could do at least some of the communication without any locks at all, using a lock-free data structure. This is worth experimenting with, but we’ll have to wait for the benchmarks to show whether this is the best idea.

Somewhat orthogonal to message passing performance, I hope to be able to implement more of the communication system in Rust itself. When the communication system was first written, we had to put almost all of it in the runtime library that was written in C++. Rust is significantly more powerful now, and it seems like it would not take too much more work to write the communication system almost entirely in Rust.