live_runtime.rs 10.5 KB
Newer Older
1
2
3
4
5
// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

use std::sync::Arc;

6
use anyhow::{Result, anyhow};
7
8
9
10
11
12
13
use dashmap::DashMap;
use dynamo_kv_router::config::KvRouterConfig;
use tokio::sync::{Notify, Semaphore, mpsc};
use tokio::task::JoinSet;
use tokio::time::Instant;
use tokio_util::sync::CancellationToken;

14
use crate::common::protocols::{DirectRequest, FpmPublisher, MockEngineArgs, OutputSignal};
15
use crate::loadgen::WorkloadDriver;
16
use crate::replay::{ReplayPrefillLoadEstimator, ReplayRouterMode, TraceSimulationReport};
17
18
use crate::scheduler::{AdmissionEvent, EngineScheduler, SchedulerHandle};

19
use super::ReplayRouter;
20
21
22
23
24
25
26
use super::demux::run_demux;
use super::state::{
    LiveReplayMode, LiveRuntimeStats, SharedLiveRuntimeStats, WorkloadDispatchState, now_ms,
    record_arrival,
};
use super::task::{RequestTaskContext, run_request_task, wait_for_workload_progress};

27
28
pub(super) struct LiveRuntime {
    pending: std::collections::VecDeque<DirectRequest>,
29
30
    senders: Arc<[mpsc::UnboundedSender<DirectRequest>]>,
    schedulers: Vec<EngineScheduler>,
31
    output_rx: mpsc::UnboundedReceiver<Vec<OutputSignal>>,
32
33
34
35
36
37
38
39
    admission_rx: mpsc::UnboundedReceiver<AdmissionEvent>,
    cancel_token: CancellationToken,
    start: Instant,
    mode: LiveReplayMode,
    router: Arc<ReplayRouter>,
}

impl LiveRuntime {
40
41
    /// Build the shared router, worker schedulers, and demux inputs for one live replay run.
    pub(super) fn new(
42
43
        args: MockEngineArgs,
        router_config: Option<KvRouterConfig>,
44
        prefill_load_estimator: Option<ReplayPrefillLoadEstimator>,
45
        pending: std::collections::VecDeque<DirectRequest>,
46
47
48
49
50
        num_workers: usize,
        mode: LiveReplayMode,
        router_mode: ReplayRouterMode,
    ) -> Result<Self> {
        let cancel_token = CancellationToken::new();
51
        let (output_tx, output_rx) = mpsc::unbounded_channel::<Vec<OutputSignal>>();
52
53
54
55
56
        let (admission_tx, admission_rx) = mpsc::unbounded_channel();
        let router = Arc::new(ReplayRouter::new(
            router_mode,
            &args,
            router_config,
57
            prefill_load_estimator,
58
59
60
61
62
63
64
65
66
67
68
69
70
            num_workers,
        ));
        let mut schedulers = Vec::with_capacity(num_workers);
        let mut senders = Vec::with_capacity(num_workers);

        for worker_idx in 0..num_workers {
            let scheduler = EngineScheduler::new_with_admission(
                args.clone(),
                0,
                Some(output_tx.clone()),
                router.sink(worker_idx as _),
                Some(cancel_token.clone()),
                Some(admission_tx.clone()),
71
                FpmPublisher::default(),
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
            );
            senders.push(scheduler.request_sender());
            schedulers.push(scheduler);
        }
        Ok(Self {
            pending,
            senders: Arc::from(senders),
            schedulers,
            output_rx,
            admission_rx,
            cancel_token,
            start: Instant::now(),
            mode,
            router,
        })
    }

89
90
    /// Replay a finite queue of requests and return the final trace report plus debug stats.
    pub(super) async fn run(mut self) -> Result<(TraceSimulationReport, LiveRuntimeStats)> {
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
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
        let requests = Arc::new(DashMap::with_capacity(self.pending.len()));
        let stats = Arc::new(SharedLiveRuntimeStats::default());
        let (arrival_tx, arrival_rx) = mpsc::unbounded_channel();
        let demux_requests = Arc::clone(&requests);
        let start = self.start;
        let router = Arc::clone(&self.router);
        let senders = Arc::clone(&self.senders);
        let output_rx = self.output_rx;
        let admission_rx = self.admission_rx;
        let demux_stats = Arc::clone(&stats);
        let demux_router = Arc::clone(&router);
        let demux_task = tokio::spawn(async move {
            run_demux(
                start,
                arrival_rx,
                admission_rx,
                output_rx,
                demux_requests,
                demux_router,
                demux_stats,
            )
            .await
        });
        let mut tasks = JoinSet::new();
        let task_ctx = RequestTaskContext {
            senders,
            router: Arc::clone(&self.router),
            requests: Arc::clone(&requests),
            stats: Arc::clone(&stats),
            workload: None,
        };

        match self.mode {
            LiveReplayMode::Trace => {
                while let Some(request) = self.pending.pop_front() {
                    let arrival_ms = request.arrival_timestamp_ms.unwrap_or(0.0);
                    let deadline =
                        start + tokio::time::Duration::from_secs_f64(arrival_ms / 1000.0);
                    tokio::time::sleep_until(deadline).await;
                    record_arrival(&arrival_tx, &request, arrival_ms)?;
                    tasks.spawn(run_request_task(task_ctx.clone(), request, None));
                }
            }
            LiveReplayMode::Concurrency { max_in_flight } => {
                let semaphore = Arc::new(Semaphore::new(max_in_flight));
                while let Some(request) = self.pending.pop_front() {
                    let permit = semaphore
                        .clone()
                        .acquire_owned()
                        .await
                        .map_err(|_| anyhow!("online replay concurrency semaphore closed"))?;
                    record_arrival(&arrival_tx, &request, now_ms(start))?;
                    tasks.spawn(run_request_task(task_ctx.clone(), request, Some(permit)));
                }
            }
        }

        while let Some(result) = tasks.join_next().await {
            result.map_err(|e| anyhow!("online replay request task failed: {e}"))??;
        }

        drop(arrival_tx);
        self.cancel_token.cancel();
        self.schedulers.clear();

        let report = demux_task
            .await
            .map_err(|e| anyhow!("online replay demux task failed: {e}"))?;
        router.shutdown().await?;
        Ok((report, stats.snapshot()))
    }

163
164
    /// Drive a multi-turn workload driver until it is drained and all spawned request tasks finish.
    pub(super) async fn run_workload(
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
        mut self,
        driver: WorkloadDriver,
        total_turns: usize,
    ) -> Result<(TraceSimulationReport, LiveRuntimeStats)> {
        let requests = Arc::new(DashMap::with_capacity(total_turns.max(1)));
        let stats = Arc::new(SharedLiveRuntimeStats::default());
        let (arrival_tx, arrival_rx) = mpsc::unbounded_channel();
        let demux_requests = Arc::clone(&requests);
        let start = self.start;
        let router = Arc::clone(&self.router);
        let senders = Arc::clone(&self.senders);
        let output_rx = self.output_rx;
        let admission_rx = self.admission_rx;
        let demux_stats = Arc::clone(&stats);
        let demux_router = Arc::clone(&router);
        let demux_task = tokio::spawn(async move {
            run_demux(
                start,
                arrival_rx,
                admission_rx,
                output_rx,
                demux_requests,
                demux_router,
                demux_stats,
            )
            .await
        });
        let workload = Arc::new(WorkloadDispatchState {
            driver: std::sync::Mutex::new(driver),
            wakeup: Notify::new(),
            start,
        });
        let mut tasks = JoinSet::new();
        let task_ctx = RequestTaskContext {
            senders,
            router: Arc::clone(&self.router),
            requests: Arc::clone(&requests),
            stats: Arc::clone(&stats),
            workload: Some(Arc::clone(&workload)),
        };
        let semaphore = match self.mode {
            LiveReplayMode::Trace => None,
            LiveReplayMode::Concurrency { max_in_flight } => {
                Some(Arc::new(Semaphore::new(max_in_flight)))
            }
        };

        loop {
            let now = now_ms(start);
            let dispatch_limit = match &semaphore {
                Some(semaphore) => semaphore.available_permits(),
                None => usize::MAX,
            };

            if dispatch_limit > 0 {
                let ready_turns = workload
                    .driver
                    .lock()
                    .unwrap()
                    .pop_ready(now, dispatch_limit);
                if !ready_turns.is_empty() {
                    for ready_turn in ready_turns {
                        let permit = match &semaphore {
                            Some(semaphore) => {
                                Some(semaphore.clone().try_acquire_owned().map_err(|_| {
                                    anyhow!(
                                        "online replay concurrency semaphore unexpectedly closed"
                                    )
                                })?)
                            }
                            None => None,
                        };
                        let arrival_at_ms = match self.mode {
                            LiveReplayMode::Trace => ready_turn.scheduled_ready_at_ms,
                            LiveReplayMode::Concurrency { .. } => now_ms(start),
                        };
                        record_arrival(&arrival_tx, &ready_turn.request, arrival_at_ms)?;
                        tasks.spawn(run_request_task(
                            task_ctx.clone(),
                            ready_turn.request,
                            permit,
                        ));
                    }
                    continue;
                }
            }

            let wake = workload.wakeup.notified();
            tokio::pin!(wake);
            let (is_drained, next_ready_ms) = {
                let mut driver = workload.driver.lock().unwrap();
                (driver.is_drained(), driver.next_ready_time_ms())
            };
            if is_drained {
                break;
            }

            wait_for_workload_progress(
                self.mode,
                semaphore.as_deref(),
                next_ready_ms,
                start,
                wake.as_mut(),
            )
            .await;
        }

        while let Some(result) = tasks.join_next().await {
            result.map_err(|e| anyhow!("online replay request task failed: {e}"))??;
        }

        drop(arrival_tx);
        self.cancel_token.cancel();
        self.schedulers.clear();

        let report = demux_task
            .await
            .map_err(|e| anyhow!("online replay demux task failed: {e}"))?;
        router.shutdown().await?;
        Ok((report, stats.snapshot()))
    }
}