Web Services with Rust Part 1: Exploring Hyper

Datetime:2017-04-17 05:27:14         Topic: Web Service  Rust          Share        Original >>
Here to See The Original Article!!!

A while ago I have turned my head towards Rust for developing Web Services. Rust’s design in my opinion hits a sweet spot between ease of development and runtime efficiency, especially doing away with a garbage collector.

Recently sophisticated support for future-based , async , and reactive programming has been added to the Rust ecosystem; putting together in the Tokio project some of the most intriguing designs I have come across so far.

In addition, the hyper Web server is being refactored to natively work with the async and reactive mechanics provided by Tokio. This work is very closely aligned with the work on Tokio, making for a very promising HTTP toolkit roadmap.

All in all, this feels a lot like the more lightweight, less bloated alternative to all I like about Akka-HTTP that I have been hoping for for quite some time now.

As sugar on top there are several very interesting developments regarding related topics, such as lock free data structures (also see Lock-freedom without garbage collection ) and Actors .

In this series of blog postings we’ll explore how to build Web services with hyper, focusing primarily on non functional aspects such as using thread pools with hyper, shared state,

Minimal, Single-Threaded Server (No Shared State, No I/O)

The most simple server to run based on hyper is a single-threaded server that does not involve shared state or blocking calls. We’ll just use the echo server example from the hyper repository.

To get some rough idea about the performance impact of our forthcoming modifications we’ll use a simple testing setup using a 4 core 2.4 GHz, 8GB RAM, 1BGbps bare metal server offered by Packet and Stormforger to put substantial stress on the Web servers. Test server and load generators are both located in continental Europe.

The first test runs with the example server capped around 10k req/s, running out of TCP connections. I added some CPU-bound work to the server, simulating, for example, template rendering activities you might have in a real Web server and had to experiment with the load tests setting to get to a point where I could drive the server to the limit in a controlled way.

The work simulation looks like this

fn cpu_intensive_work() -> String {
    let mut y = "X".to_string();
    for x in 0..100 {
        y = format!("Value: {}", x);
    let address = Address {
        street: "10 Downing Street".to_owned(),
        city: y.to_owned(),

    let j = serde_json::to_string(&address).unwrap();
    return j;

The server code I compiled as a statically linked executable for Linux.

The Stormforger load test looks like this:


    duration: 60,
    rate: 8.0,         // clients per second to launch
    max_clients: 500,

  cluster: { sizing: "small", },

definition.session("Load Test 1", function(session) {

    context.get("/data", { tag: "root" });


With that I was able to get about 6000 req/s out of the server before seeing the 99th percentile latency degrading. modifying the simulated work showed corresponding variations in req/s numbers so I am now sure I am not hitting any unwanted capacity limits.

Now that we have a baseline, let’s see what happens if we involve multiple cores.

Minimal Multi-Threaded Server (No Shared State, No I/O)

—- —-