chore: add benchmarks for no-op submission (#144)

Co-authored-by: ollie-etl <Oliver Bunting@etlsystems.com>

.github/workflows/ci.yml

@@ -27,6 +27,14 @@ jobs:
     steps:
       - run: exit 0
 
+  bench:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - name: Install Rust
+        run: rustup update stable
+      - run: cargo bench --no-run
+
   check:
     runs-on: ubuntu-latest
     steps:
@@ -59,7 +67,6 @@ jobs:
         run: rustup update stable
       - run: cargo test --doc
 
-
   fmt:
     runs-on: ubuntu-latest
     steps:

Cargo.toml

@@ -28,6 +28,19 @@ bytes = { version = "1.0", optional = true }
 [dev-dependencies]
 tempfile = "3.2.0"
 tokio-test = "0.4.2"
+iai = "0.1.1"
+futures = "0.3.25"
+criterion = "0.4.0"
 
 [package.metadata.docs.rs]
 all-features = true
+
+[[bench]]
+name = "lai_no_op"
+path = "benches/lai/no_op.rs"
+harness = false
+
+[[bench]]
+name = "criterion_no_op"
+path = "benches/criterion/no_op.rs"
+harness = false

benches/criterion/no_op.rs

@@ -0,0 +1,79 @@
+use criterion::{
+    criterion_group, criterion_main, BenchmarkId, Criterion, SamplingMode, Throughput,
+};
+use std::time::{Duration, Instant};
+
+use futures::stream::{self, StreamExt};
+
+#[derive(Clone)]
+struct Options {
+    iterations: usize,
+    concurrency: usize,
+    sq_size: usize,
+    cq_size: usize,
+}
+
+impl Default for Options {
+    fn default() -> Self {
+        Self {
+            iterations: 100000,
+            concurrency: 1,
+            sq_size: 128,
+            cq_size: 256,
+        }
+    }
+}
+
+fn run_no_ops(opts: &Options, count: u64) -> Duration {
+    let mut ring_opts = tokio_uring::uring_builder();
+    ring_opts
+        .setup_cqsize(opts.cq_size as _)
+        // .setup_sqpoll(10)
+        // .setup_sqpoll_cpu(1)
+        ;
+
+    let mut m = Duration::ZERO;
+
+    // Run the required number of iterations
+    for _ in 0..count {
+        m += tokio_uring::builder()
+            .entries(opts.sq_size as _)
+            .uring_builder(&ring_opts)
+            .start(async move {
+                let start = Instant::now();
+                stream::iter(0..opts.iterations)
+                    .for_each_concurrent(Some(opts.concurrency), |_| async move {
+                        tokio_uring::no_op().await.unwrap();
+                    })
+                    .await;
+                start.elapsed()
+            })
+    }
+    m
+}
+
+fn bench(c: &mut Criterion) {
+    let mut group = c.benchmark_group("no_op");
+    let mut opts = Options::default();
+    for concurrency in [1, 32, 64, 256].iter() {
+        opts.concurrency = *concurrency;
+
+        // We perform long running benchmarks: this is the best mode
+        group.sampling_mode(SamplingMode::Flat);
+
+        group.throughput(Throughput::Elements(opts.iterations as u64));
+        group.bench_with_input(
+            BenchmarkId::from_parameter(concurrency),
+            &opts,
+            |b, opts| {
+                // Custom iterator used because we don't expose access to runtime,
+                // which is required to do async benchmarking with criterion
+                b.iter_custom(move |iter| run_no_ops(opts, iter));
+            },
+        );
+    }
+    group.finish();
+}
+
+criterion_group!(benches, bench);
+criterion_main!(benches);

benches/lai/no_op.rs

@@ -0,0 +1,83 @@
+use futures::stream::{self, StreamExt};
+use iai::black_box;
+
+#[derive(Clone)]
+struct Options {
+    iterations: usize,
+    concurrency: usize,
+    sq_size: usize,
+    cq_size: usize,
+}
+
+impl Default for Options {
+    fn default() -> Self {
+        Self {
+            iterations: 100000,
+            concurrency: 1,
+            sq_size: 64,
+            cq_size: 256,
+        }
+    }
+}
+
+fn runtime_only() -> Result<(), Box<dyn std::error::Error>> {
+    let opts = Options::default();
+    let mut ring_opts = tokio_uring::uring_builder();
+    ring_opts
+        .setup_cqsize(opts.cq_size as _)
+        // .setup_sqpoll(10)
+        // .setup_sqpoll_cpu(1)
+        ;
+
+    tokio_uring::builder()
+        .entries(opts.sq_size as _)
+        .uring_builder(&ring_opts)
+        .start(async move { black_box(Ok(())) })
+}
+
+fn run_no_ops(opts: Options) -> Result<(), Box<dyn std::error::Error>> {
+    let mut ring_opts = tokio_uring::uring_builder();
+    ring_opts
+        .setup_cqsize(opts.cq_size as _)
+        // .setup_sqpoll(10)
+        // .setup_sqpoll_cpu(1)
+        ;
+
+    tokio_uring::builder()
+        .entries(opts.sq_size as _)
+        .uring_builder(&ring_opts)
+        .start(async move {
+            stream::iter(0..opts.iterations)
+                .for_each_concurrent(Some(opts.concurrency), |_| async move {
+                    tokio_uring::no_op().await.unwrap();
+                })
+                .await;
+            Ok(())
+        })
+}
+
+// This provides a baseline for estimating op overhead on top of this
+fn no_op_x1() -> Result<(), Box<dyn std::error::Error>> {
+    let opts = Options::default();
+    run_no_ops(black_box(opts))
+}
+
+fn no_op_x32() -> Result<(), Box<dyn std::error::Error>> {
+    let mut opts = Options::default();
+    opts.concurrency = 32;
+    run_no_ops(black_box(opts))
+}
+
+fn no_op_x64() -> Result<(), Box<dyn std::error::Error>> {
+    let mut opts = Options::default();
+    opts.concurrency = 64;
+    run_no_ops(black_box(opts))
+}
+
+fn no_op_x256() -> Result<(), Box<dyn std::error::Error>> {
+    let mut opts = Options::default();
+    opts.concurrency = 256;
+    run_no_ops(black_box(opts))
+}
+
+iai::main!(runtime_only, no_op_x1, no_op_x32, no_op_x64, no_op_x256);