chore: fix criterion benchmarks

Follow up to tokio-rs#151.

benches/criterion/no_op.rs

@@ -1,25 +1,56 @@
-use criterion::{
-    criterion_group, criterion_main, BenchmarkId, Criterion, SamplingMode, Throughput,
-};
+/// Benchmark the overhead of the uring no_op operation.
+///
+/// It should be understood that a single call to even the no_op operation is relatively expensive
+/// even though the kernel is not being asked to do anything except copy the user_data from the
+/// submission queue entry to a completion queue entry.
+///
+/// The cost comes from the application creating writing to the SQE, then just before going idle,
+/// making the system call to notify the kernel that the submission queue tail has been changed by
+/// the application, then being awoken at the mio level by the kernel that something about the
+/// uring's fd has become ready, then running through the tokio_uring's completion ring handler and
+/// finding the CQE that trigger's a lookup into the slab that gets a future's awake function
+/// called, that makes a task runnable, that then allows the Tokio runtime to schedule the task
+/// that finally gets to see the no_op operation completed.
+///
+/// When run in isolation, that is to say when then concurrency parameter is 1, this round trip
+/// takes between 3.2 and 1.5 *micro* seconds on a 64 bit Linux machine - depending on the CPU
+/// speed and the cache sizes.
+///
+/// When run with enough concurrency, say 1000 tasks each doing the same thing, it can be seen the
+/// actual overhead for the tokio_uring no_op call and wait comes down to between 310 and 220ns,
+/// again depending on CPU and cache details. Given that a trivial yield by a Tokio task itself
+/// takes generally 100ns, this overhead for a uring operation is not much.
+///
+/// It just takes a lot of concurrent work going through the uring interface to see that benefit.
+use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion, SamplingMode};
 use std::time::{Duration, Instant};
 
 use tokio::task::JoinSet;
 
 #[derive(Clone)]
 struct Options {
-    iterations: usize,
-    concurrency: usize,
+    // Variable, per benchmark.
+    concurrency: u64,
+
+    // Constants, same for all benchmarks.
     sq_size: usize,
     cq_size: usize,
 }
 
 impl Default for Options {
     fn default() -> Self {
         Self {
-            iterations: 100000,
+            // Default to 1, but each benchmark actually sets the concurrenty level.
             concurrency: 1,
-            sq_size: 128,
-            cq_size: 256,
+
+            // We run up to 10K concurrent tasks in this benchmark, but optimize for 1K, so create
+            // an input ring of the same size. We understand 1,000 would be rounded up to 1024, so
+            // use that size explicitly anyway.
+            sq_size: 1024,
+
+            // The completion ring is going to be double the submission ring by default, so this
+            // wouldn't be necessary unless we want to experiment with larger multiples.
+            cq_size: 2 * 1024,
         }
     }
 }
@@ -28,44 +59,52 @@ 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 _);
 
-    let mut m = Duration::ZERO;
+    tokio_uring::builder()
+        .entries(opts.sq_size as _)
+        .uring_builder(&ring_opts)
+        .start(async move {
+            let mut js = JoinSet::new();
 
-    // 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 mut js = JoinSet::new();
+            // Prepare the number of concurrent tasks this benchmark calls for.
+            // They will each be blocked until the current task yields the thread
+            // when it calls await on the first join_next below.
+            //
+            // Within each task, run through the number of no_op calls in series based on the
+            // benchmark count (accounting for the number of tasks that are being spawned).
 
-                for _ in 0..opts.iterations {
-                    js.spawn_local(tokio_uring::no_op());
-                }
+            for _ in 0..opts.concurrency {
+                let opts = opts.clone();
+                js.spawn_local(async move {
+                    // Run the required number of iterations (per the concurrency amount)
+                    for _ in 0..(count / opts.concurrency) {
+                        let _ = black_box(tokio_uring::no_op().await);
+                    }
+                });
+            }
 
-                let start = Instant::now();
+            // Measure from the moment we start waiting for the futures.
+            let start = Instant::now();
 
-                while let Some(res) = js.join_next().await {
-                    res.unwrap().unwrap();
-                }
+            while let Some(res) = js.join_next().await {
+                res.unwrap();
+            }
 
-                start.elapsed()
-            })
-    }
-    m
+            // Return the elapsed time.
+            start.elapsed()
+        })
 }
 
 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() {
+    for concurrency in [1, 2, 5, 10, 100, 1000, 10_000].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),
+            BenchmarkId::new("concurrent tasks", concurrency),
             &opts,
             |b, opts| {
                 // Custom iterator used because we don't expose access to runtime,