cherenkov/server/
pacer.rs

1//! Wall-time pacing within the prefill capacity reserved at load.
2
3pub(super) struct ChunkPacer {
4    floor: usize,
5    ceiling: usize,
6    /// Zero uses the configured ceiling for every chunk.
7    target_seconds: f64,
8    tokens: usize,
9    contended: bool,
10}
11
12impl ChunkPacer {
13    pub(super) fn new(ceiling: usize, target_seconds: f64) -> Self {
14        let floor = ceiling.min(128);
15
16        Self {
17            floor,
18            ceiling,
19            target_seconds,
20            tokens: if target_seconds == 0.0 {
21                ceiling
22            } else {
23                floor
24            },
25            contended: false,
26        }
27    }
28
29    /// A new contention period starts small. Work already on the GPU still
30    /// finishes its current chunk before another request can run.
31    pub(super) fn quantum(&mut self, contended: bool) -> usize {
32        if self.target_seconds == 0.0 {
33            return self.ceiling;
34        }
35
36        if contended && !self.contended {
37            self.tokens = self.floor;
38        }
39
40        self.contended = contended;
41
42        if contended { self.tokens } else { self.ceiling }
43    }
44
45    /// Adjust the target from a successful engine chunk. The caller excludes
46    /// short tails, capacity-limited chunks, and the small-row path.
47    pub(super) fn observe(&mut self, tokens: usize, seconds: f64) {
48        if !self.contended || tokens == 0 || !seconds.is_finite() || seconds <= 0.0 {
49            return;
50        }
51
52        let scaled = tokens as f64 * self.target_seconds / seconds;
53        let damped = scaled.clamp(self.tokens as f64 * 0.5, self.tokens as f64 * 2.0);
54
55        self.tokens = (damped as usize).clamp(self.floor, self.ceiling);
56    }
57
58    /// Current contended chunk target, before memory and boundary limits.
59    pub(super) fn tokens(&self) -> usize {
60        self.tokens
61    }
62}
63
64#[cfg(test)]
65#[path = "../../tests/unit/server/pacer.rs"]
66mod tests;
cherenkov/server/pacer.rs

cherenkov/server/
pacer.rs
