Add benchmark for full stack request lifecycle (#1298)

* add benchmark for full stack request lifecycle

* add direct service benchmarks

* fix newline

* add cloneable service benchmarks

* remove cloneable bench experiments + cargo fmt

Co-authored-by: Yuki Okushi <huyuumi.dev@gmail.com>

Cargo.toml

@@ -99,6 +99,7 @@ env_logger = "0.6"
 serde_derive = "1.0"
 brotli2 = "0.3.2"
 flate2 = "1.0.13"
+criterion = "0.3"
 
 [profile.release]
 lto = true
@@ -116,3 +117,11 @@ actix-session = { path = "actix-session" }
 actix-files = { path = "actix-files" }
 actix-multipart = { path = "actix-multipart" }
 awc = { path = "awc" }
+
+[[bench]]
+name = "server"
+harness = false
+
+[[bench]]
+name = "service"
+harness = false

benches/server.rs

@@ -0,0 +1,64 @@
+use actix_web::{test, web, App, HttpResponse};
+use awc::Client;
+use criterion::{criterion_group, criterion_main, Criterion};
+use futures::future::join_all;
+
+const STR: &str = "Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World \
+                   Hello World Hello World Hello World Hello World Hello World";
+
+// benchmark sending all requests at the same time
+fn bench_async_burst(c: &mut Criterion) {
+    let srv = test::start(|| {
+        App::new()
+            .service(web::resource("/").route(web::to(|| HttpResponse::Ok().body(STR))))
+    });
+
+    // We are using System here, since Runtime requires preinitialized tokio
+    // Maybe add to actix_rt docs
+    let url = srv.url("/");
+    let mut rt = actix_rt::System::new("test");
+
+    c.bench_function("get_body_async_burst", move |b| {
+        b.iter_custom(|iters| {
+            let client = Client::new().get(url.clone()).freeze().unwrap();
+
+            let start = std::time::Instant::now();
+            // benchmark body
+            let resps = rt.block_on(async move {
+                let burst = (0..iters).map(|_| client.send());
+                join_all(burst).await
+            });
+            let elapsed = start.elapsed();
+
+            // if there are failed requests that might be an issue
+            let failed = resps.iter().filter(|r| r.is_err()).count();
+            if failed > 0 {
+                eprintln!("failed {} requests (might be bench timeout)", failed);
+            };
+
+            elapsed
+        })
+    });
+}
+
+criterion_group!(server_benches, bench_async_burst);
+criterion_main!(server_benches);

benches/service.rs

@@ -0,0 +1,108 @@
+use actix_service::Service;
+use actix_web::dev::{ServiceRequest, ServiceResponse};
+use actix_web::{web, App, Error, HttpResponse};
+use criterion::{criterion_main, Criterion};
+use std::cell::RefCell;
+use std::rc::Rc;
+
+use actix_web::test::{init_service, ok_service, TestRequest};
+
+/// Criterion Benchmark for async Service
+/// Should be used from within criterion group:
+/// ```rust,ignore
+/// let mut criterion: ::criterion::Criterion<_> =
+///     ::criterion::Criterion::default().configure_from_args();
+/// bench_async_service(&mut criterion, ok_service(), "async_service_direct");
+/// ```
+///
+/// Usable for benching Service wrappers:
+/// Using minimum service code implementation we first measure
+/// time to run minimum service, then measure time with wrapper.
+///
+/// Sample output
+/// async_service_direct    time:   [1.0908 us 1.1656 us 1.2613 us]
+pub fn bench_async_service<S>(c: &mut Criterion, srv: S, name: &str)
+where
+    S: Service<Request = ServiceRequest, Response = ServiceResponse, Error = Error>
+        + 'static,
+{
+    let mut rt = actix_rt::System::new("test");
+    let srv = Rc::new(RefCell::new(srv));
+
+    let req = TestRequest::default().to_srv_request();
+    assert!(rt
+        .block_on(srv.borrow_mut().call(req))
+        .unwrap()
+        .status()
+        .is_success());
+
+    // start benchmark loops
+    c.bench_function(name, move |b| {
+        b.iter_custom(|iters| {
+            let srv = srv.clone();
+            // exclude request generation, it appears it takes significant time vs call (3us vs 1us)
+            let reqs: Vec<_> = (0..iters)
+                .map(|_| TestRequest::default().to_srv_request())
+                .collect();
+            let start = std::time::Instant::now();
+            // benchmark body
+            rt.block_on(async move {
+                for req in reqs {
+                    srv.borrow_mut().call(req).await.unwrap();
+                }
+            });
+            let elapsed = start.elapsed();
+            // check that at least first request succeeded
+            elapsed
+        })
+    });
+}
+
+async fn index(req: ServiceRequest) -> Result<ServiceResponse, Error> {
+    Ok(req.into_response(HttpResponse::Ok().finish()))
+}
+
+// Benchmark basic WebService directly
+// this approach is usable for benching WebService, though it adds some time to direct service call:
+// Sample results on MacBook Pro '14
+// time:   [2.0724 us 2.1345 us 2.2074 us]
+fn async_web_service(c: &mut Criterion) {
+    let mut rt = actix_rt::System::new("test");
+    let srv = Rc::new(RefCell::new(rt.block_on(init_service(
+        App::new().service(web::service("/").finish(index)),
+    ))));
+
+    let req = TestRequest::get().uri("/").to_request();
+    assert!(rt
+        .block_on(srv.borrow_mut().call(req))
+        .unwrap()
+        .status()
+        .is_success());
+
+    // start benchmark loops
+    c.bench_function("async_web_service_direct", move |b| {
+        b.iter_custom(|iters| {
+            let srv = srv.clone();
+            let reqs = (0..iters).map(|_| TestRequest::get().uri("/").to_request());
+
+            let start = std::time::Instant::now();
+            // benchmark body
+            rt.block_on(async move {
+                for req in reqs {
+                    srv.borrow_mut().call(req).await.unwrap();
+                }
+            });
+            let elapsed = start.elapsed();
+            // check that at least first request succeeded
+            elapsed
+        })
+    });
+}
+
+pub fn service_benches() {
+    let mut criterion: ::criterion::Criterion<_> =
+        ::criterion::Criterion::default().configure_from_args();
+    bench_async_service(&mut criterion, ok_service(), "async_service_direct");
+    async_web_service(&mut criterion);
+}
+criterion_main!(service_benches);