pool_lat/
pool_lat.rs

1//! Pool dispatch-latency probe (scratch diagnostic, not a product).
2//!
3//! Times `Pool::run` with an empty closure: the pure cost of waking every
4//! worker and collecting the latch. Decode issues one dispatch per matvec
5//! (~200/token), so this cost is paid ~200 times per token — compare it
6//! against the per-worker work of a typical MLP row-slice.
7//!
8//! Usage: cargo run --release --example pool_lat
9
10use cortiq_engine::pool::Pool;
11use std::time::Instant;
12
13fn main() {
14    for nt in [2usize, 4, 6, 8, 10] {
15        let pool = Pool::new(nt);
16        let noop: &(dyn Fn(usize, usize) + Sync) = &|_w, _n| {};
17        // Warm: first dispatch spawns/parks the workers.
18        for _ in 0..100 {
19            pool.run(noop);
20        }
21        let iters = 20_000;
22        let t0 = Instant::now();
23        for _ in 0..iters {
24            pool.run(noop);
25        }
26        let el = t0.elapsed().as_secs_f64();
27        let per = el / iters as f64 * 1e6;
28        println!(
29            "threads={nt:2}  {per:7.2} us/dispatch  -> {:6.2} ms/token at 200 matvecs",
30            per * 200.0 / 1e3
31        );
32    }
33}
pool_lat/pool_lat.rs

pool_lat/
pool_lat.rs
