threadteam 0.2.0

Scoped CPU thread-team placement token: a Pool and a non-Clone Exec with fork, for_ranges, external and broadcast; no Rayon type in the public API.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135

//! Dispatch latency and idle CPU against the park delay (`idle_spin`) and the
//! gap between sections, for a team of 6 participants.
//!
//! For each (spin, gap) cell a fresh team is built, then the leader repeats
//! "wait `gap`, dispatch a trivial section to all participants". Reported per
//! cell:
//!
//! * `lat`: median time from the leader's call to the *last* participant
//!   starting its share (first-dispatch latency after the gap);
//! * `cores`: CPU time of the process other than the leader thread, divided by
//!   wall time, over the whole cell: the cost of members that spin or wake
//!   (the members' idle CPU in units of cores; 5 = all five members spinning).
//!
//! The leader waits by sleeping when `gap >= 1 ms` and by busy-waiting below
//! (a sleep cannot be shorter than that); the busy-wait is the leader's own
//! thread and is excluded from `cores`.
//!
//! `cargo bench --bench spin_grid -- [cell_seconds]`

use std::sync::atomic::{AtomicU64, Ordering};
use std::time::{Duration, Instant};
use threadteam::{raise_current_thread_qos, Pool, Qos};

fn cpu_ns(clock: libc::clockid_t) -> u64 {
    let mut ts = libc::timespec {
        tv_sec: 0,
        tv_nsec: 0,
    };
    unsafe { libc::clock_gettime(clock, &mut ts) };
    ts.tv_sec as u64 * 1_000_000_000 + ts.tv_nsec as u64
}

fn wait(gap: Duration) {
    if gap >= Duration::from_millis(1) {
        std::thread::sleep(gap);
    } else {
        let t = Instant::now();
        while t.elapsed() < gap {
            std::hint::spin_loop();
        }
    }
}

fn median(v: &mut [f64]) -> f64 {
    v.sort_by(f64::total_cmp);
    v[v.len() / 2]
}

fn cell(spin: Duration, gap: Duration, seconds: f64) -> (f64, f64, f64) {
    let pool = Pool::builder()
        .threads(6)
        .qos(Qos::UserInitiated)
        .idle_spin(spin)
        .acquire_timeout(Duration::from_secs(30))
        .build()
        .unwrap();
    let team = pool.into_team(6).unwrap();
    let mut exec = team.exec(6);
    let base = Instant::now();
    let starts: Vec<AtomicU64> = (0..6).map(|_| AtomicU64::new(0)).collect();
    let one = |exec: &mut threadteam::Exec<'_>| -> f64 {
        for s in &starts {
            s.store(0, Ordering::Relaxed);
        }
        let t0 = base.elapsed().as_nanos() as u64;
        exec.reborrow()
            .broadcast(|i| starts[i].store(base.elapsed().as_nanos() as u64, Ordering::Release))
            .unwrap();
        starts
            .iter()
            .map(|s| s.load(Ordering::Acquire).saturating_sub(t0))
            .max()
            .unwrap() as f64
    };
    for _ in 0..5 {
        wait(gap);
        one(&mut exec);
    }
    let n = ((seconds / (gap.as_secs_f64() + 20e-6)) as usize).clamp(20, 100_000);
    let (p0, l0, w0) = (
        cpu_ns(libc::CLOCK_PROCESS_CPUTIME_ID),
        cpu_ns(libc::CLOCK_THREAD_CPUTIME_ID),
        Instant::now(),
    );
    let mut lat = Vec::with_capacity(n);
    for _ in 0..n {
        wait(gap);
        lat.push(one(&mut exec));
    }
    let wall = w0.elapsed().as_nanos() as f64;
    let others = (cpu_ns(libc::CLOCK_PROCESS_CPUTIME_ID) - p0) as f64
        - (cpu_ns(libc::CLOCK_THREAD_CPUTIME_ID) - l0) as f64;
    let mut sorted = lat.clone();
    let med = median(&mut sorted);
    let p90 = sorted[sorted.len() * 9 / 10];
    (med * 1e-3, p90 * 1e-3, others / wall)
}

fn main() {
    let seconds: f64 = std::env::args()
        .skip(1)
        .find_map(|a| a.parse().ok())
        .unwrap_or(1.0);
    raise_current_thread_qos(Qos::UserInitiated);
    let spins = [
        Duration::ZERO,
        Duration::from_micros(50),
        Duration::from_micros(200),
        Duration::from_millis(1),
        Duration::from_millis(10),
        Duration::from_millis(200),
    ];
    let gaps = [
        Duration::from_micros(10),
        Duration::from_micros(100),
        Duration::from_millis(1),
        Duration::from_millis(10),
        Duration::from_millis(100),
    ];
    println!("team of 6, QoS UserInitiated, cell time about {seconds} s");
    println!("each cell: median / p90 dispatch latency in us, members' CPU in cores");
    print!("{:>10}", "spin\\gap");
    for g in gaps {
        print!("{:>26}", format!("{g:?}"));
    }
    println!();
    for s in spins {
        print!("{:>10}", format!("{s:?}"));
        for g in gaps {
            let (med, p90, cores) = cell(s, g, seconds);
            print!("{:>26}", format!("{med:6.1} /{p90:6.1} us {cores:4.2}c"));
        }
        println!();
    }
}