offload.rs 58.9 KB
Newer Older
1
2
3
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

4
5
6
7
8
//! # Offload Manager
//! The offload manager is responsible for handling all block transfers between different cache levels.
//!
//! ## Offloading
//! Offloading is the process of moving blocks to a cache level further away from the device.
Ryan Olson's avatar
Ryan Olson committed
9
//! When blocks are registered (via [`ManagedBlockPool::register_blocks`]), they are automatically sent to the offload manager.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
//! Due to limited bandwidth, the offload manager must prioritize which offloads to perform.
//! This is indicated by the `priority` parameter to [`OffloadManager::offload`].
//! When a offload request is received, the offload manager will enqueue it into a priority queue.
//! This priority queue is keyed by the `priority` parameter, where blocks with lower priority values are processed first.
//! Within the same priority, blocks that were sent to the offload manager earlier are processed first.
//!
//! ## Onboarding
//! Onboarding is the process of moving blocks to a cache level closer to the device.
//! All onboardings are manually triggered through the [`OffloadManager::onboard`] method.
//!
//! ## Transfer Managers
//! The offload manager uses two transfer managers to handle the offloading and onboarding of blocks.
//!
//! The [`CudaTransferManager`] is responsible for transfers between the device and host.
//! The [`DiskTransferManager`] is responsible for transfers from host to disk and disk to device.
//!
//! ## Worker Threads
//! The offload manager uses two kinds of worker threads to handle the offloading and onboarding of blocks.
//!
//! The [`OffloadManager::offload_worker`] is responsible for offloading blocks.
//! The [`OffloadManager::onboard_worker`] is responsible for onboarding blocks.
//!
//! The kind of offloads/onboards they perform is dictated by the source and target arguments
33
//! of the [`OffloadManager::offload_worker`] and [`OffloadManager::onboard_worker`] methods.
34

Ryan Olson's avatar
Ryan Olson committed
35
use super::block::{
36
37
    BlockError, BlockMetadata, BlockState, ImmutableBlock, MutableBlock,
    locality::LocalityProvider, transfer::TransferContext,
Ryan Olson's avatar
Ryan Olson committed
38
};
39
use super::metrics::{BlockManagerMetrics, PoolMetrics};
Ryan Olson's avatar
Ryan Olson committed
40
use super::pool::{BlockPool, BlockPoolError};
41
use super::storage::{Cuda, Storage};
Ryan Olson's avatar
Ryan Olson committed
42
use super::{DeviceStorage, DiskStorage, PinnedStorage};
43
44
45
46
use nixl_sys::Agent as NixlAgent;
use std::sync::Arc;
use tokio::runtime::Handle;
use tokio::sync::{
47
    Mutex,
48
    mpsc::{self, error::TryRecvError},
49
    oneshot,
50
};
51
use tokio_util::sync::CancellationToken;
52
53
54
55
56
57
58

use anyhow::Result;
use std::any::Any;

use std::collections::BTreeSet;

mod pending;
59
pub mod request;
60

Ryan Olson's avatar
Ryan Olson committed
61
use pending::{LocalTransferManager, PendingTransfer, TransferBatcher, TransferManager};
62
use request::{BlockResult, OffloadRequest, OffloadRequestKey, OnboardRequest};
63

64
65
use dynamo_runtime::utils::task::CriticalTaskExecutionHandle;

66
67
const MAX_CONCURRENT_TRANSFERS: usize = 4;
const MAX_TRANSFER_BATCH_SIZE: usize = 16;
68
69

/// The offload manager handles all block transfers between different cache levels.
Ryan Olson's avatar
Ryan Olson committed
70
pub struct OffloadManager<Locality: LocalityProvider, Metadata: BlockMetadata> {
71
    // Handles to the device, host, and disk pools.
Ryan Olson's avatar
Ryan Olson committed
72
73
74
    disk: Option<Arc<dyn BlockPool<DiskStorage, Locality, Metadata>>>,
    host: Option<Arc<dyn BlockPool<PinnedStorage, Locality, Metadata>>>,
    device: Option<Arc<dyn BlockPool<DeviceStorage, Locality, Metadata>>>,
75

76
    /// Queue of offloading requests.
Ryan Olson's avatar
Ryan Olson committed
77
78
    device_offload_tx: mpsc::UnboundedSender<OffloadRequest<DeviceStorage, Locality, Metadata>>,
    host_offload_tx: mpsc::UnboundedSender<OffloadRequest<PinnedStorage, Locality, Metadata>>,
79
80

    /// Queue of pending onboarding requests.
Ryan Olson's avatar
Ryan Olson committed
81
82
83
84
    host_onboard_tx:
        mpsc::UnboundedSender<OnboardRequest<PinnedStorage, DeviceStorage, Locality, Metadata>>,
    disk_onboard_tx:
        mpsc::UnboundedSender<OnboardRequest<DiskStorage, DeviceStorage, Locality, Metadata>>,
85
86
87

    /// An incrementing counter for offloaded blocks. Within the same priority, blocks with lower tick values are processed first.
    tick: Arc<Mutex<u64>>,
88
89
}

Ryan Olson's avatar
Ryan Olson committed
90
91
92
impl<Locality: LocalityProvider + 'static, Metadata: BlockMetadata>
    OffloadManager<Locality, Metadata>
{
93
    pub fn new(
Ryan Olson's avatar
Ryan Olson committed
94
95
96
        disk: Option<Arc<dyn BlockPool<DiskStorage, Locality, Metadata>>>,
        host: Option<Arc<dyn BlockPool<PinnedStorage, Locality, Metadata>>>,
        device: Option<Arc<dyn BlockPool<DeviceStorage, Locality, Metadata>>>,
97
98
        nixl_agent: Arc<Option<NixlAgent>>,
        async_rt_handle: Handle,
99
        metrics: Arc<BlockManagerMetrics>,
100
        cancellation_token: CancellationToken,
101
    ) -> Result<Arc<Self>> {
102
103
104
105
106
        let (device_offload_tx, device_offload_rx) = mpsc::unbounded_channel();
        let (host_offload_tx, host_offload_rx) = mpsc::unbounded_channel();

        let (host_onboard_tx, host_onboard_rx) = mpsc::unbounded_channel();
        let (disk_onboard_tx, disk_onboard_rx) = mpsc::unbounded_channel();
107
108

        let this = Arc::new(Self {
109
            disk,
110
            host,
111
112
113
114
115
            device,
            device_offload_tx,
            host_offload_tx,
            host_onboard_tx,
            disk_onboard_tx,
116
117
118
            tick: Arc::new(Mutex::new(0)),
        });

119
        let cuda_ctx = Cuda::device_or_create(0)?;
120

121
122
123
124
        // We want cuda offloads to happen in parallel with host onboards, so we need to use a different stream.
        let device_offload_transfer_ctx = Arc::new(TransferContext::new(
            nixl_agent.clone(),
            cuda_ctx.new_stream()?,
125
            async_rt_handle.clone(),
126
        ));
127

Ryan Olson's avatar
Ryan Olson committed
128
129
130
131
        let device_metrics = metrics.pool("device");
        let host_metrics = metrics.pool("host");
        let disk_metrics = metrics.pool("disk");

132
        // Device -> Host offload
133
134
135
136
137
        let device_to_host_task = OffloadManager::offload_worker(
            this.device.clone(),
            this.host.clone(),
            device_offload_rx,
            Arc::new(TransferBatcher::new(
Ryan Olson's avatar
Ryan Olson committed
138
                LocalTransferManager::new(
139
140
                    device_offload_transfer_ctx,
                    MAX_CONCURRENT_TRANSFERS,
141
                    &async_rt_handle,
142
                    cancellation_token.clone(),
Ryan Olson's avatar
Ryan Olson committed
143
144
                    device_metrics.clone(),
                    "offload_bw".to_string(),
145
                )?,
146
147
148
149
                MAX_TRANSFER_BATCH_SIZE,
                &async_rt_handle,
                cancellation_token.clone(),
            )),
Ryan Olson's avatar
Ryan Olson committed
150
            device_metrics.clone(),
151
152
153
154
155
156
157
158
159
            cancellation_token.clone(),
        );
        CriticalTaskExecutionHandle::new_with_runtime(
            |_| device_to_host_task,
            cancellation_token.clone(),
            "Device -> Host offload worker",
            &async_rt_handle,
        )?
        .detach();
160

161
162
163
        let transfer_ctx = Arc::new(TransferContext::new(
            nixl_agent.clone(),
            cuda_ctx.new_stream()?,
164
            async_rt_handle.clone(),
165
        ));
166

167
        // Host -> Disk offload
168
169
170
171
172
        let host_to_disk_task = OffloadManager::offload_worker(
            this.host.clone(),
            this.disk.clone(),
            host_offload_rx,
            Arc::new(TransferBatcher::new(
Ryan Olson's avatar
Ryan Olson committed
173
                LocalTransferManager::new(
174
175
176
177
                    transfer_ctx.clone(),
                    MAX_CONCURRENT_TRANSFERS,
                    &async_rt_handle,
                    cancellation_token.clone(),
Ryan Olson's avatar
Ryan Olson committed
178
179
                    host_metrics.clone(),
                    "offload_bw".to_string(),
180
                )?,
181
182
183
184
                MAX_TRANSFER_BATCH_SIZE,
                &async_rt_handle,
                cancellation_token.clone(),
            )),
Ryan Olson's avatar
Ryan Olson committed
185
            host_metrics.clone(),
186
187
188
189
190
191
192
193
194
            cancellation_token.clone(),
        );
        CriticalTaskExecutionHandle::new_with_runtime(
            |_| host_to_disk_task,
            cancellation_token.clone(),
            "Host -> Disk offload worker",
            &async_rt_handle,
        )?
        .detach();
195

196
        // Host -> Device onboarding
197
198
199
200
201
        let host_to_device_task = OffloadManager::onboard_worker(
            this.host.clone(),
            this.device.clone(),
            host_onboard_rx,
            Arc::new(TransferBatcher::new(
Ryan Olson's avatar
Ryan Olson committed
202
                LocalTransferManager::new(
203
204
                    transfer_ctx.clone(),
                    MAX_CONCURRENT_TRANSFERS,
205
                    &async_rt_handle,
206
                    cancellation_token.clone(),
Ryan Olson's avatar
Ryan Olson committed
207
208
                    host_metrics.clone(),
                    "onboard_bw".to_string(),
209
                )?,
210
211
212
213
                MAX_TRANSFER_BATCH_SIZE,
                &async_rt_handle,
                cancellation_token.clone(),
            )),
Ryan Olson's avatar
Ryan Olson committed
214
            host_metrics.clone(),
215
216
217
218
219
220
221
222
223
            cancellation_token.clone(),
        );
        CriticalTaskExecutionHandle::new_with_runtime(
            |_| host_to_device_task,
            cancellation_token.clone(),
            "Host -> Device onboarding worker",
            &async_rt_handle,
        )?
        .detach();
224

225
        // Disk -> Device onboarding
226
227
228
229
230
        let disk_to_device_task = OffloadManager::onboard_worker(
            this.disk.clone(),
            this.device.clone(),
            disk_onboard_rx,
            Arc::new(TransferBatcher::new(
Ryan Olson's avatar
Ryan Olson committed
231
                LocalTransferManager::new(
232
233
234
235
                    transfer_ctx.clone(),
                    MAX_CONCURRENT_TRANSFERS,
                    &async_rt_handle,
                    cancellation_token.clone(),
Ryan Olson's avatar
Ryan Olson committed
236
237
                    disk_metrics.clone(),
                    "onboard_bw".to_string(),
238
                )?,
239
240
241
242
                MAX_TRANSFER_BATCH_SIZE,
                &async_rt_handle,
                cancellation_token.clone(),
            )),
Ryan Olson's avatar
Ryan Olson committed
243
            disk_metrics.clone(),
244
245
246
247
248
249
250
251
252
            cancellation_token.clone(),
        );
        CriticalTaskExecutionHandle::new_with_runtime(
            |_| disk_to_device_task,
            cancellation_token.clone(),
            "Disk -> Device onboarding worker",
            &async_rt_handle,
        )?
        .detach();
253

254
        Ok(this)
255
    }
256

257
    async fn offload_worker<Source: Storage, Target: Storage>(
Ryan Olson's avatar
Ryan Olson committed
258
259
260
261
        source_pool: Option<Arc<dyn BlockPool<Source, Locality, Metadata>>>,
        target_pool: Option<Arc<dyn BlockPool<Target, Locality, Metadata>>>,
        mut offload_rx: mpsc::UnboundedReceiver<OffloadRequest<Source, Locality, Metadata>>,
        transfer_manager: Arc<dyn TransferManager<Source, Target, Locality, Metadata>>,
262
        pool_metrics: Arc<PoolMetrics>,
263
        cancellation_token: CancellationToken,
264
    ) -> Result<()> {
265
        if source_pool.is_none() || target_pool.is_none() {
266
267
268
            return Ok(());
        }

269
270
        let source_pool = source_pool.as_ref().unwrap();
        let target_pool = target_pool.as_ref().unwrap();
271

272
        let mut queue = BTreeSet::new();
273
274

        loop {
275
276
277
278
            if cancellation_token.is_cancelled() {
                return Ok(());
            }

279
            // Try to check the offload queue.
280
281
282
283
            loop {
                match offload_rx.try_recv() {
                    Ok(request) => {
                        queue.insert(request);
284
                        pool_metrics.gauge("offload_queue_size").inc();
285
286
287
288
                    }
                    Err(TryRecvError::Empty) => {
                        break;
                    }
289
                    Err(e) => return Err(e.into()),
290
291
                }
            }
292
293

            // If there is a request, process it.
294
            if let Some(request) = queue.pop_first() {
295
                pool_metrics.gauge("offload_queue_size").dec();
296
297
                // Try to upgrade the block to a strong reference.
                let block = match request.block.upgrade() {
Ryan Olson's avatar
Ryan Olson committed
298
                    Some(block) => Some(ImmutableBlock::new(block)),
299
                    // If unable to upgrade, the block may have been moved to the inactive pool.
300
                    None => source_pool
301
302
                        .match_sequence_hashes(vec![request.sequence_hash].as_slice())
                        .await?
Ryan Olson's avatar
Ryan Olson committed
303
                        .pop(),
304
305
                };

306
                // If we've found the block, offload it.
307
                if let Some(block) = block {
308
309
                    // If the block is already in the target, don't offload it.
                    if let Ok(blocks) = target_pool
Ryan Olson's avatar
Ryan Olson committed
310
311
                        .match_sequence_hashes(vec![request.sequence_hash].as_slice())
                        .await
312
                        && !blocks.is_empty()
313
                    {
314
                        continue;
315
316
                    }

317
                    let target_block = 'target_block: {
318
319
320
321
                        if let Ok(blocks) = target_pool.allocate_blocks(1).await
                            && let Some(block) = blocks.into_iter().next()
                        {
                            break 'target_block Some(block);
322
                        }
323

324
325
326
                        tracing::warn!(
                            "Target pool full. Skipping offload. This should only ever happen with very small pool sizes."
                        );
327
                        None
328
329
                    };

330
                    if let Some(target_block) = target_block {
331
                        pool_metrics.counter("offload_processed").inc();
Ryan Olson's avatar
Ryan Olson committed
332
333
334
335
                        tracing::debug!(
                            "Offloading block with sequence hash {} to target pool.",
                            request.sequence_hash
                        );
336
                        transfer_manager
337
                            .enqueue_transfer(PendingTransfer::new(
338
                                vec![block],
339
                                vec![target_block],
340
                                None,
341
                                target_pool.clone(),
342
343
344
345
346
                            ))
                            .await?;
                    }
                }
            } else {
347
                // Await the next request.
348
349
350
351
                tokio::select! {
                    _ = cancellation_token.cancelled() => return Ok(()),
                    Some(request) = offload_rx.recv() => {
                        queue.insert(request);
352
                        pool_metrics.gauge("offload_queue_size").inc();
353
                    }
354
                }
355
356
357
358
            }
        }
    }

359
    async fn onboard_worker<Source: Storage, Target: Storage>(
Ryan Olson's avatar
Ryan Olson committed
360
361
362
363
        source_pool: Option<Arc<dyn BlockPool<Source, Locality, Metadata>>>,
        target_pool: Option<Arc<dyn BlockPool<Target, Locality, Metadata>>>,
        mut onboard_rx: mpsc::UnboundedReceiver<OnboardRequest<Source, Target, Locality, Metadata>>,
        transfer_manager: Arc<dyn TransferManager<Source, Target, Locality, Metadata>>,
364
        pool_metrics: Arc<PoolMetrics>,
365
        cancellation_token: CancellationToken,
366
    ) -> Result<()> {
367
        if source_pool.is_none() || target_pool.is_none() {
368
369
370
            return Ok(());
        }

371
        let target_pool = target_pool.as_ref().unwrap();
372
373
374
375
        loop {
            tokio::select! {
                _ = cancellation_token.cancelled() => return Ok::<(), anyhow::Error>(()),
                Some(request) = onboard_rx.recv() => {
376
377
378
379
380

                    pool_metrics
                        .gauge("onboard_queue_size")
                        .set(onboard_rx.len() as i64);

381
                    // Try to allocate blocks on the device.
Ryan Olson's avatar
Ryan Olson committed
382
383
384
385
386
387
388
389
390
                    let target_blocks = if let Some(targets) = request.targets {
                        targets
                    } else {
                            match target_pool.allocate_blocks(request.blocks.len()).await {
                            Ok(blocks) => blocks,
                            Err(err) => {
                                let _ = request.response_tx.send(Err(err));
                                continue;
                            }
391
392
                        }
                    };
393

394
395
396
397
                    pool_metrics
                        .counter("onboard_processed")
                        .inc_by(request.blocks.len() as u64);

Ryan Olson's avatar
Ryan Olson committed
398
                    tracing::debug!("Onboarding {} blocks to target pool.", request.blocks.len());
399
400
401

                    transfer_manager
                        .enqueue_transfer(PendingTransfer::new(
Ryan Olson's avatar
Ryan Olson committed
402
                            request.blocks,
403
404
405
406
407
408
409
                            target_blocks,
                            Some(request.response_tx),
                            target_pool.clone(),
                        ))
                        .await?;

                    Ok::<(), anyhow::Error>(())
410
                }
411
            }?;
412
413
414
415
416
        }
    }

    pub async fn offload<S: Storage>(
        &self,
Ryan Olson's avatar
Ryan Olson committed
417
        block: &ImmutableBlock<S, Locality, Metadata>,
418
419
420
        priority: u64,
    ) -> core::result::Result<(), BlockPoolError> {
        match block.state() {
421
            BlockState::Registered(_, _) => {}
422
423
424
425
426
427
            _ => {
                return Err(BlockPoolError::BlockError(BlockError::InvalidState(
                    "Block is not registered.".to_string(),
                )));
            }
        }
428
429
430
431
432
433
434
435
436
437

        let mut tick = self.tick.lock().await;
        let key = OffloadRequestKey {
            priority,
            timestamp: *tick,
        };
        // Increment a counter for each block. Within the same priority, blocks with lower counter values are processed first.
        *tick += 1;
        drop(tick);

438
439
440
441
442
443
        // This can get called by all pools, regardless of whether or not they have a place to offload to.
        // Because of this, we need to check the block type here.
        let any_block = block as &dyn Any;

        // TODO: What's the performance penalty of this runtime type-checking?
        if let Some(device_block) =
Ryan Olson's avatar
Ryan Olson committed
444
            any_block.downcast_ref::<ImmutableBlock<DeviceStorage, Locality, Metadata>>()
445
        {
446
447
448
449
            // The host pool doesn't exist, so we can't offload to it.
            if self.device_offload_tx.is_closed() {
                return Ok(());
            }
450
451
452

            let request = OffloadRequest {
                block: Arc::downgrade(device_block.mutable_block()),
Ryan Olson's avatar
Ryan Olson committed
453
                sequence_hash: device_block.sequence_hash(),
454
455
456
                key,
            };

457
458
            self.device_offload_tx.send(request).unwrap();
        } else if let Some(host_block) =
Ryan Olson's avatar
Ryan Olson committed
459
            any_block.downcast_ref::<ImmutableBlock<PinnedStorage, Locality, Metadata>>()
460
461
462
463
464
465
466
467
        {
            // The disk pool doesn't exist, so we can't offload to it.
            if self.host_offload_tx.is_closed() {
                return Ok(());
            }

            let request = OffloadRequest {
                block: Arc::downgrade(host_block.mutable_block()),
Ryan Olson's avatar
Ryan Olson committed
468
                sequence_hash: host_block.sequence_hash(),
469
470
471
472
                key,
            };

            self.host_offload_tx.send(request).unwrap();
473
474
475
476
477
        }

        Ok(())
    }

Ryan Olson's avatar
Ryan Olson committed
478
    pub fn onboard<S: Storage>(
479
        &self,
Ryan Olson's avatar
Ryan Olson committed
480
481
482
483
        blocks: Vec<ImmutableBlock<S, Locality, Metadata>>,
        targets: Option<Vec<MutableBlock<DeviceStorage, Locality, Metadata>>>,
    ) -> oneshot::Receiver<BlockResult<DeviceStorage, Locality, Metadata>> {
        let (tx, rx) = oneshot::channel();
484
485
        for block in &blocks {
            match block.state() {
486
                BlockState::Registered(_, _) => {}
487
                _ => {
Ryan Olson's avatar
Ryan Olson committed
488
                    tx.send(Err(BlockPoolError::BlockError(BlockError::InvalidState(
489
                        "Block is not registered.".to_string(),
Ryan Olson's avatar
Ryan Olson committed
490
491
492
                    ))))
                    .unwrap();
                    return rx;
493
494
495
496
                }
            }
        }

497
498
499
500
501
502
503
504
        if let Some(targets) = targets.as_ref()
            && targets.len() != blocks.len()
        {
            tx.send(Err(BlockPoolError::BlockError(BlockError::Other(
                anyhow::anyhow!("Number of targets does not match number of blocks."),
            ))))
            .unwrap();
            return rx;
505
506
        }

Ryan Olson's avatar
Ryan Olson committed
507
508
509
510
        if blocks.is_empty() {
            tx.send(Ok(vec![])).unwrap();
            return rx;
        }
511

512
513
514
515
        let any_block = blocks.first().unwrap() as &dyn Any;

        // TODO: This is really ugly.
        if any_block
Ryan Olson's avatar
Ryan Olson committed
516
            .downcast_ref::<ImmutableBlock<PinnedStorage, Locality, Metadata>>()
517
518
519
520
521
522
            .is_some()
        {
            let host_blocks = blocks
                .iter()
                .map(|b| {
                    (b as &dyn Any)
Ryan Olson's avatar
Ryan Olson committed
523
                        .downcast_ref::<ImmutableBlock<PinnedStorage, Locality, Metadata>>()
524
525
526
527
528
                        .unwrap()
                        .clone()
                })
                .collect();

Ryan Olson's avatar
Ryan Olson committed
529
530
531
532
533
534
535
536
            if let Err(e) = self
                .host_onboard_tx
                .send(OnboardRequest::new(host_blocks, tx, targets))
            {
                e.0.response_tx
                    .send(Err(BlockPoolError::ProgressEngineShutdown))
                    .unwrap();
            }
537
        } else if any_block
Ryan Olson's avatar
Ryan Olson committed
538
            .downcast_ref::<ImmutableBlock<DiskStorage, Locality, Metadata>>()
539
540
541
542
543
544
            .is_some()
        {
            let disk_blocks = blocks
                .iter()
                .map(|b| {
                    (b as &dyn Any)
Ryan Olson's avatar
Ryan Olson committed
545
                        .downcast_ref::<ImmutableBlock<DiskStorage, Locality, Metadata>>()
546
547
548
549
550
                        .unwrap()
                        .clone()
                })
                .collect();

Ryan Olson's avatar
Ryan Olson committed
551
552
553
554
555
556
557
558
            if let Err(e) = self
                .disk_onboard_tx
                .send(OnboardRequest::new(disk_blocks, tx, targets))
            {
                e.0.response_tx
                    .send(Err(BlockPoolError::ProgressEngineShutdown))
                    .unwrap();
            }
559
        } else {
Ryan Olson's avatar
Ryan Olson committed
560
            tx.send(Err(BlockPoolError::BlockError(BlockError::Other(
561
                anyhow::anyhow!("Block type not supported for onboarding."),
Ryan Olson's avatar
Ryan Olson committed
562
563
            ))))
            .unwrap();
564
565
        }

Ryan Olson's avatar
Ryan Olson committed
566
        rx
567
568
569
570
    }
}

#[cfg(all(test, feature = "testing-cuda"))]
Ryan Olson's avatar
Ryan Olson committed
571
mod tests {
572
573
574
    use super::*;

    use crate::block_manager::{
575
        LayoutConfig, NixlRegisterableStorage,
576
        block::{
577
            BasicMetadata, BlockDataExt, BlockDataProvider, Blocks, MutableBlock, locality::Local,
578
        },
579
        layout::{FullyContiguous, LayerSeparate, LayoutType, nixl::NixlLayout},
Ryan Olson's avatar
Ryan Olson committed
580
        pool::{BlockRegistrationDuplicationSetting, ManagedBlockPool},
581
        storage::{
582
            DeviceAllocator, DeviceStorage, DiskAllocator, DiskStorage, PinnedAllocator,
Ryan Olson's avatar
Ryan Olson committed
583
            PinnedStorage, StorageAllocator, StorageType,
584
585
        },
    };
586
    use crate::tokens::{TokenBlockSequence, Tokens};
587
    use nixl_sys::{MemoryRegion, NixlDescriptor};
588

589
    use aligned_vec::avec;
590
    use cudarc::runtime::sys::{cudaMemcpy, cudaMemcpyKind, cudaMemset};
591
    use prometheus::Registry;
Ryan Olson's avatar
Ryan Olson committed
592
    use rstest::*;
593
    use std::fs::File;
594
    use std::io::{Read, Seek, SeekFrom, Write};
595
596
    use std::mem::ManuallyDrop;
    use std::os::unix::io::FromRawFd;
597
598

    const BLOCK_SIZE: usize = 4;
599
    const NUM_LAYERS: usize = 8;
600

Ryan Olson's avatar
Ryan Olson committed
601
602
603
    type DevicePool = Option<Arc<dyn BlockPool<DeviceStorage, Local, BasicMetadata>>>;
    type HostPool = Option<Arc<dyn BlockPool<PinnedStorage, Local, BasicMetadata>>>;
    type DiskPool = Option<Arc<dyn BlockPool<DiskStorage, Local, BasicMetadata>>>;
604
605
606
607
608

    lazy_static::lazy_static! {
        static ref NIXL_AGENT: Arc<Option<NixlAgent>> = {
            let agent = NixlAgent::new("offload-manager").unwrap();
            let (_, ucx_params) = agent.get_plugin_params("UCX").unwrap();
Ryan Olson's avatar
Ryan Olson committed
609
            let (_, gds_mt_params) = agent.get_plugin_params("GDS_MT").unwrap();
610
            let (_, posix_params) = agent.get_plugin_params("POSIX").unwrap();
611
            agent.create_backend("UCX", &ucx_params).unwrap();
Ryan Olson's avatar
Ryan Olson committed
612
            agent.create_backend("GDS_MT", &gds_mt_params).unwrap();
613
            agent.create_backend("POSIX", &posix_params).unwrap();
614
615
616
            Arc::new(Some(agent))
        };
    }
617

Ryan Olson's avatar
Ryan Olson committed
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
    fn build_layout<S: Storage + NixlRegisterableStorage>(
        config: LayoutConfig,
        layout_type: LayoutType,
        agent: &NixlAgent,
        allocator: &dyn StorageAllocator<S>,
        duplication_setting: BlockRegistrationDuplicationSetting,
    ) -> Result<Arc<dyn BlockPool<S, Local, BasicMetadata>>> {
        match layout_type {
            LayoutType::FullyContiguous => {
                let mut pool_layout = FullyContiguous::allocate(config.clone(), allocator)?;
                pool_layout.nixl_register(agent, None)?;
                let blocks = Blocks::new(pool_layout, 42, 0)?.into_blocks()?;
                Ok(Arc::new(
                    ManagedBlockPool::builder()
                        .blocks(blocks)
                        .default_duplication_setting(duplication_setting)
                        .build()?,
                ))
            }
            LayoutType::LayerSeparate { outer_contiguous } => {
                let mut pool_layout =
                    LayerSeparate::allocate(config.clone(), allocator, outer_contiguous)?;
                pool_layout.nixl_register(agent, None)?;
                let blocks = Blocks::new(pool_layout, 42, 0)?.into_blocks()?;
                Ok(Arc::new(
                    ManagedBlockPool::builder()
                        .blocks(blocks)
                        .default_duplication_setting(duplication_setting)
                        .build()?,
                ))
            }
        }
    }

    #[allow(clippy::type_complexity)]
    fn build_pools(
654
655
        device_blocks: usize,
        host_blocks: Option<usize>,
656
        disk_blocks: Option<usize>,
657
        inner_dim: Option<usize>,
658
    ) -> Result<(
Ryan Olson's avatar
Ryan Olson committed
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
        Arc<OffloadManager<Local, BasicMetadata>>,
        DevicePool,
        HostPool,
        DiskPool,
    )> {
        build_pools_with_layout(
            device_blocks,
            host_blocks,
            disk_blocks,
            inner_dim,
            LayoutType::FullyContiguous,
            BlockRegistrationDuplicationSetting::Disabled,
        )
    }

    #[allow(clippy::type_complexity)]
    pub fn build_pools_with_layout(
        device_blocks: usize,
        host_blocks: Option<usize>,
        disk_blocks: Option<usize>,
        inner_dim: Option<usize>,
        layout_type: LayoutType,
        duplication_setting: BlockRegistrationDuplicationSetting,
    ) -> Result<(
        Arc<OffloadManager<Local, BasicMetadata>>,
684
685
686
687
        DevicePool,
        HostPool,
        DiskPool,
    )> {
688
689
        let mut config = LayoutConfig {
            num_blocks: device_blocks,
690
            num_layers: NUM_LAYERS,
691
            outer_dim: 1,
692
            page_size: BLOCK_SIZE,
693
            inner_dim: inner_dim.unwrap_or(1024),
694
            alignment: 1,
Ryan Olson's avatar
Ryan Olson committed
695
            dtype_width_bytes: 2,
696
697
        };

698
699
700
        let agent_arc = NIXL_AGENT.clone();
        let agent = agent_arc.as_ref().as_ref().unwrap();

Ryan Olson's avatar
Ryan Olson committed
701
702
703
704
705
706
707
        let device_pool = Some(build_layout(
            config.clone(),
            layout_type,
            agent,
            &DeviceAllocator::default(),
            duplication_setting,
        )?);
708
709
710

        let host_pool = if let Some(host_blocks) = host_blocks {
            config.num_blocks = host_blocks;
Ryan Olson's avatar
Ryan Olson committed
711
712
713
714
715
716
717
            Some(build_layout(
                config.clone(),
                layout_type,
                agent,
                &PinnedAllocator::default(),
                duplication_setting,
            )?)
718
        } else {
719
            None
720
721
        };

722
723
        let disk_pool = if let Some(disk_blocks) = disk_blocks {
            config.num_blocks = disk_blocks;
Ryan Olson's avatar
Ryan Olson committed
724
725
726
727
728
729
730
            Some(build_layout(
                config,
                layout_type,
                agent,
                &DiskAllocator,
                duplication_setting,
            )?)
731
        } else {
732
            None
733
        };
734

735
736
737
738
739
740
741
742
        let async_rt_handle = Handle::current();

        let manager = OffloadManager::new(
            disk_pool.clone(),
            host_pool.clone(),
            device_pool.clone(),
            agent_arc,
            async_rt_handle,
743
            BlockManagerMetrics::new(&Arc::new(Registry::new()))?,
744
            CancellationToken::new(),
745
746
747
        )?;

        Ok((manager, device_pool, host_pool, disk_pool))
748
749
750
    }

    /// Create a block in the 'RESET' state.
Ryan Olson's avatar
Ryan Olson committed
751
    #[expect(dead_code)]
752
    async fn get_block<S: Storage, Metadata: BlockMetadata>(
Ryan Olson's avatar
Ryan Olson committed
753
754
755
756
        pool: &Arc<dyn BlockPool<S, Local, Metadata>>,
    ) -> Result<MutableBlock<S, Local, Metadata>> {
        let mut blocks = pool.allocate_blocks(1).await?;
        Ok(blocks.pop().unwrap())
757
758
759
760
    }

    /// Create a block in the 'COMPLETED' state.
    async fn completed_block<S: Storage, Metadata: BlockMetadata>(
Ryan Olson's avatar
Ryan Olson committed
761
        pool: &Arc<dyn BlockPool<S, Local, Metadata>>,
762
        tokens: [u32; BLOCK_SIZE],
Ryan Olson's avatar
Ryan Olson committed
763
764
765
766
767
768
769
770
    ) -> Result<MutableBlock<S, Local, Metadata>> {
        let mut block = pool
            .allocate_blocks(1)
            .await?
            .into_iter()
            .next()
            .ok_or(anyhow::anyhow!("Failed to allocate block"))?;

771
772
773
774
775
776
777
778
        block.init_sequence(42)?;
        for token in tokens {
            block.add_token(token)?;
        }
        block.commit()?;
        Ok(block)
    }

779
    fn populate_block<S: Storage + NixlDescriptor>(
780
        block: &impl BlockDataProvider<StorageType = S>,
Ryan Olson's avatar
Ryan Olson committed
781
        start_value: u8,
782
    ) -> Result<()> {
Ryan Olson's avatar
Ryan Olson committed
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
        let block_data = block.block_data();

        let mut value = start_value;

        for layer_idx in 0..block_data.num_layers() {
            for outer_idx in 0..block_data.num_outer_dims() {
                let layer_view = block_data.layer_view(layer_idx, outer_idx)?;
                match block_data.storage_type() {
                    StorageType::Device(_) | StorageType::Pinned => unsafe {
                        cudaMemset(
                            layer_view.as_ptr() as *mut std::ffi::c_void,
                            value as i32,
                            layer_view.size(),
                        )
                        .result()?;
                    },
                    StorageType::Disk(_) => {
                        let nixl_desc = layer_view.as_nixl_descriptor();
                        let mut file: ManuallyDrop<File>;
                        let data = avec![[4096] | value; layer_view.size()];

                        unsafe {
                            file =
                                ManuallyDrop::new(File::from_raw_fd(nixl_desc.device_id() as i32));
                            file.seek(SeekFrom::Start(nixl_desc.as_ptr() as u64))?;
                        }
                        file.write_all(&data)?;
                        file.sync_all()?;
                        file.flush()?;
                    }
                    _ => panic!(),
814
815
                }
            }
Ryan Olson's avatar
Ryan Olson committed
816
817

            value += 1;
818
        }
819

820
821
822
        Ok(())
    }

823
824
    fn get_block_contents<S: Storage + NixlDescriptor>(
        block: &impl BlockDataProvider<StorageType = S>,
Ryan Olson's avatar
Ryan Olson committed
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
    ) -> Result<Vec<Vec<u8>>> {
        let block_data = block.block_data();

        let mut contents: Vec<Vec<u8>> = Vec::new();

        for layer_idx in 0..block_data.num_layers() {
            for outer_idx in 0..block_data.num_outer_dims() {
                let layer_view = block_data.layer_view(layer_idx, outer_idx)?;
                match block_data.storage_type() {
                    StorageType::Device(_) => unsafe {
                        let mut buffer = vec![0_u8; layer_view.size()];

                        cudaMemcpy(
                            buffer.as_mut_ptr() as *mut std::ffi::c_void,
                            layer_view.as_ptr() as *const std::ffi::c_void,
                            layer_view.size(),
                            cudaMemcpyKind::cudaMemcpyDeviceToHost,
                        )
                        .result()?;

                        contents.push(buffer);
                    },
                    StorageType::Pinned => unsafe {
                        contents.push(
                            std::slice::from_raw_parts(layer_view.as_ptr(), layer_view.size())
                                .to_vec(),
                        );
                    },
                    StorageType::Disk(_) => {
                        let nixl_desc = layer_view.as_nixl_descriptor();
                        let mut file: ManuallyDrop<File>;
                        let mut aligned = avec![[4096] | 0; layer_view.size()];

                        unsafe {
                            file =
                                ManuallyDrop::new(File::from_raw_fd(nixl_desc.device_id() as i32));
                            file.seek(SeekFrom::Start(nixl_desc.as_ptr() as u64))?;
                        }
                        file.read_exact(&mut aligned)?;
                        contents.push(aligned.to_vec());
                    }
                    _ => anyhow::bail!("Unsupported storage type."),
867
868
                }
            }
869
870
        }

Ryan Olson's avatar
Ryan Olson committed
871
        Ok(contents)
872
873
    }

874
    fn check_block_contents(
875
876
        block1: &impl BlockDataProvider<StorageType = impl Storage + NixlDescriptor>,
        block2: &impl BlockDataProvider<StorageType = impl Storage + NixlDescriptor>,
Ryan Olson's avatar
Ryan Olson committed
877
        start_value: u8,
878
    ) -> Result<()> {
879
880
        let contents1 = get_block_contents(block1)?;
        let contents2 = get_block_contents(block2)?;
881

Ryan Olson's avatar
Ryan Olson committed
882
883
884
885
886
887
888
889
890
        assert_eq!(contents1.len(), contents2.len());

        let mut value = start_value;

        for (layer1_vec, layer2_vec) in contents1.iter().zip(contents2.iter()) {
            for (c1_value, c2_value) in layer1_vec.iter().zip(layer2_vec.iter()) {
                if c1_value != c2_value || c1_value != &value {
                    panic!("{} != {} != {}", c1_value, c2_value, value);
                }
891
            }
Ryan Olson's avatar
Ryan Olson committed
892
            value += 1;
893
        }
894
895
896
897
898
        Ok(())
    }

    #[tokio::test]
    async fn test_offload_invalid_blocks() -> Result<()> {
899
        let (offload_manager, device_pool, _, _) = build_pools(4, Some(4), None, None)?;
900

901
        let device_pool = device_pool.as_ref().unwrap();
902
903
904
905
906
907
908
909
910
911
912
913
914
915

        // Check blocks in the 'COMPLETED' state.
        let immutable_block = ImmutableBlock::new(Arc::new(
            completed_block(device_pool, [0; BLOCK_SIZE]).await?,
        ));
        assert!(matches!(
            offload_manager.offload(&immutable_block, 0).await,
            Err(BlockPoolError::BlockError(BlockError::InvalidState(_)))
        ));

        Ok(())
    }

    #[tokio::test]
Ryan Olson's avatar
Ryan Olson committed
916
917
918
919
920
921
922
923
924
925
926
927
928
    #[rstest]
    #[case(LayoutType::FullyContiguous)]
    #[case(LayoutType::LayerSeparate { outer_contiguous: true })]
    #[case(LayoutType::LayerSeparate { outer_contiguous: false })]
    async fn test_offload_registered_blocks(#[case] layout_type: LayoutType) -> Result<()> {
        let (offload_manager, device_pool, host_pool, _) = build_pools_with_layout(
            4,
            Some(4),
            None,
            None,
            layout_type,
            BlockRegistrationDuplicationSetting::Disabled,
        )?;
929

930
931
        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();
932
933
934
935
936
937
938
939
940
941
942

        // Create a block and register it with the offload manager
        let block = completed_block(device_pool, [0, 1, 2, 3]).await?;

        let immutable_device_block = device_pool
            .register_blocks(vec![block])
            .await?
            .into_iter()
            .next()
            .ok_or(anyhow::anyhow!("Failed to register block"))?;

943
        populate_block(&immutable_device_block, 42)?;
944
945
946
947
948
949
950
951
952
953
954

        // Offloads should only go to G2 (for now)
        offload_manager.offload(&immutable_device_block, 0).await?;

        // Wait for it to be processed.
        // TODO: This is a bit of a hack, and may lead to non-deterministic behavior.
        // In theory, the offload + memcpy should take much less time than this.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        // Check that the block exists in the host pool
        let host_blocks = host_pool
Ryan Olson's avatar
Ryan Olson committed
955
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
956
957
958
959
            .await?;

        assert_eq!(host_blocks.len(), 1);
        assert_eq!(
Ryan Olson's avatar
Ryan Olson committed
960
961
            host_blocks[0].sequence_hash(),
            immutable_device_block.sequence_hash()
962
963
        );

964
        check_block_contents(&immutable_device_block, &host_blocks[0], 42)?;
965
966
967
968
969
970

        Ok(())
    }

    #[tokio::test]
    async fn test_no_host_blocks_available() -> Result<()> {
971
        let (offload_manager, device_pool, host_pool, _) = build_pools(4, Some(4), None, None)?;
972

973
974
        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993

        let host_blocks = host_pool.allocate_blocks(4).await?;
        assert_eq!(host_blocks.len(), 4);

        let device_block = completed_block(device_pool, [0, 1, 2, 3]).await?;
        let immutable_device_block = device_pool
            .register_blocks(vec![device_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

        offload_manager.offload(&immutable_device_block, 0).await?;

        // Wait for offload to be processed.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        // The offload should fail gracefuly due to a lack of host blocks
        let matched_host_blocks = host_pool
Ryan Olson's avatar
Ryan Olson committed
994
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
            .await?;
        assert_eq!(matched_host_blocks.len(), 0);

        // Wait for blocks to be returned to the pool.
        drop(host_blocks);
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        // Try the offload again.
        offload_manager.offload(&immutable_device_block, 0).await?;

        // Wait for offload to be processed.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        // This time, the offload should succeed.
        let matched_host_blocks = host_pool
Ryan Olson's avatar
Ryan Olson committed
1010
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
1011
1012
1013
1014
1015
1016
1017
            .await?;
        assert_eq!(matched_host_blocks.len(), 1);

        Ok(())
    }

    #[tokio::test]
Ryan Olson's avatar
Ryan Olson committed
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
    #[rstest]
    #[case(LayoutType::FullyContiguous)]
    #[case(LayoutType::LayerSeparate { outer_contiguous: true })]
    #[case(LayoutType::LayerSeparate { outer_contiguous: false })]
    async fn test_onboard(#[case] layout_type: LayoutType) -> Result<()> {
        let (offload_manager, device_pool, host_pool, _) = build_pools_with_layout(
            4,
            Some(4),
            None,
            None,
            layout_type,
            BlockRegistrationDuplicationSetting::Disabled,
        )?;
1031

1032
1033
        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043

        // Allocate and fill a block on the host.
        let host_block = completed_block(host_pool, [0, 1, 2, 3]).await?;
        let immutable_host_block = host_pool
            .register_blocks(vec![host_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

1044
        populate_block(&immutable_host_block, 42)?;
1045
1046
1047

        // Onboard the block.
        let onboarded_blocks = offload_manager
Ryan Olson's avatar
Ryan Olson committed
1048
1049
            .onboard(vec![immutable_host_block.clone()], None)
            .await??;
1050
1051
1052
1053

        assert_eq!(onboarded_blocks.len(), 1);
        // Check that the sequence hash is the same.
        assert_eq!(
Ryan Olson's avatar
Ryan Olson committed
1054
1055
            onboarded_blocks[0].sequence_hash(),
            immutable_host_block.sequence_hash()
1056
1057
1058
1059
        );
        // Check that the block is registered.
        assert!(matches!(
            onboarded_blocks[0].state(),
1060
            BlockState::Registered(_, _)
1061
1062
        ));

1063
        check_block_contents(&immutable_host_block, &onboarded_blocks[0], 42)?;
1064
1065
1066
1067

        // Wait for the new value to show up in the device pool.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;
        let device_blocks = device_pool
Ryan Olson's avatar
Ryan Olson committed
1068
            .match_sequence_hashes(vec![onboarded_blocks[0].sequence_hash()].as_slice())
1069
1070
1071
            .await?;
        assert_eq!(device_blocks.len(), 1);
        assert_eq!(
Ryan Olson's avatar
Ryan Olson committed
1072
1073
            device_blocks[0].sequence_hash(),
            onboarded_blocks[0].sequence_hash()
1074
1075
1076
        );

        // Check that this is the same block.
1077
        check_block_contents(&immutable_host_block, &device_blocks[0], 42)?;
1078
1079
1080
1081
1082

        Ok(())
    }

    #[tokio::test]
Ryan Olson's avatar
Ryan Olson committed
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
    #[rstest]
    #[case(LayoutType::FullyContiguous)]
    #[case(LayoutType::LayerSeparate { outer_contiguous: true })]
    #[case(LayoutType::LayerSeparate { outer_contiguous: false })]
    async fn test_offload_onboard(#[case] layout_type: LayoutType) -> Result<()> {
        let (offload_manager, device_pool, host_pool, _) = build_pools_with_layout(
            4,
            Some(4),
            None,
            None,
            layout_type,
            BlockRegistrationDuplicationSetting::Disabled,
        )?;
1096

1097
1098
        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();
1099
1100
1101
1102
1103
1104
1105
1106
1107

        let device_block = completed_block(device_pool, [0, 1, 2, 3]).await?;
        let immutable_device_block = device_pool
            .register_blocks(vec![device_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

1108
        populate_block(&immutable_device_block, 42)?;
1109
1110
1111
1112
1113
1114
1115
1116
        // Offload the block to the host.
        offload_manager.offload(&immutable_device_block, 0).await?;

        // Wait for the offload to be processed.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        // Check that the block exists in the host pool.
        let immutable_host_block = host_pool
Ryan Olson's avatar
Ryan Olson committed
1117
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
1118
1119
1120
1121
1122
            .await?
            .into_iter()
            .next()
            .unwrap();

1123
        check_block_contents(&immutable_device_block, &immutable_host_block, 42)?;
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138

        // Remove the device block from the pool by dropping it and allocating more blocks.
        drop(immutable_device_block);

        // Wait for the block to be returned to the pool.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        let device_blocks = device_pool.allocate_blocks(4).await?;
        assert_eq!(device_blocks.len(), 4);

        drop(device_blocks);
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        // Check that the block is not in the device pool.
        let device_blocks = device_pool
Ryan Olson's avatar
Ryan Olson committed
1139
            .match_sequence_hashes(vec![immutable_host_block.sequence_hash()].as_slice())
1140
1141
1142
1143
1144
            .await?;
        assert_eq!(device_blocks.len(), 0);

        // Onboard the block back to the device pool.
        let onboarded_blocks = offload_manager
Ryan Olson's avatar
Ryan Olson committed
1145
1146
            .onboard(vec![immutable_host_block.clone()], None)
            .await??;
1147
1148
        assert_eq!(onboarded_blocks.len(), 1);
        assert_eq!(
Ryan Olson's avatar
Ryan Olson committed
1149
1150
            onboarded_blocks[0].sequence_hash(),
            immutable_host_block.sequence_hash()
1151
1152
1153
        );
        assert!(matches!(
            onboarded_blocks[0].state(),
1154
            BlockState::Registered(_, _)
1155
1156
        ));

1157
        check_block_contents(&immutable_host_block, &onboarded_blocks[0], 42)?;
1158
1159
1160
1161
1162
1163

        Ok(())
    }

    #[tokio::test]
    async fn test_onboard_err_handling() -> Result<()> {
1164
        let (offload_manager, device_pool, host_pool, _) = build_pools(4, Some(4), None, None)?;
1165

1166
1167
        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180

        let host_block = completed_block(host_pool, [0, 1, 2, 3]).await?;
        let immutable_host_block = host_pool
            .register_blocks(vec![host_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

        let device_blocks = device_pool.allocate_blocks(4).await?;
        assert_eq!(device_blocks.len(), 4);

        let res = offload_manager
Ryan Olson's avatar
Ryan Olson committed
1181
1182
            .onboard(vec![immutable_host_block.clone()], None)
            .await?;
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
        assert!(matches!(
            res.err().unwrap(),
            BlockPoolError::NotEnoughBlocksAvailable(_, _)
        ));

        Ok(())
    }

    #[tokio::test]
    async fn test_offload_onboard_no_host_blocks() -> Result<()> {
1193
        let (offload_manager, device_pool, _, _) = build_pools(4, None, None, None)?;
1194

1195
        let device_pool = device_pool.as_ref().unwrap();
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208

        let device_block = completed_block(device_pool, [0, 1, 2, 3]).await?;
        let immutable_device_block = device_pool
            .register_blocks(vec![device_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

        offload_manager.offload(&immutable_device_block, 0).await?;

        Ok(())
    }
1209
1210

    #[tokio::test]
Ryan Olson's avatar
Ryan Olson committed
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
    #[rstest]
    #[case(LayoutType::FullyContiguous)]
    #[case(LayoutType::LayerSeparate { outer_contiguous: true })]
    #[case(LayoutType::LayerSeparate { outer_contiguous: false })]
    async fn test_offload_disk(#[case] layout_type: LayoutType) -> Result<()> {
        let (offload_manager, _, host_pool, disk_pool) = build_pools_with_layout(
            4,
            Some(4),
            Some(4),
            None,
            layout_type,
            BlockRegistrationDuplicationSetting::Disabled,
        )?;
1224

1225
1226
        let host_pool = host_pool.as_ref().unwrap();
        let disk_pool = disk_pool.as_ref().unwrap();
1227
1228
1229
1230
1231
1232
1233
1234
1235

        let host_block = completed_block(host_pool, [0, 1, 2, 3]).await?;
        let immutable_host_block = host_pool
            .register_blocks(vec![host_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

1236
        populate_block(&immutable_host_block, 42)?;
1237
1238
1239
1240
1241
1242

        offload_manager.offload(&immutable_host_block, 0).await?;

        tokio::time::sleep(std::time::Duration::from_millis(500)).await;

        let disk_blocks = disk_pool
Ryan Olson's avatar
Ryan Olson committed
1243
            .match_sequence_hashes(vec![immutable_host_block.sequence_hash()].as_slice())
1244
1245
1246
            .await?;
        assert_eq!(disk_blocks.len(), 1);
        assert_eq!(
Ryan Olson's avatar
Ryan Olson committed
1247
1248
            disk_blocks[0].sequence_hash(),
            immutable_host_block.sequence_hash()
1249
1250
        );

1251
        check_block_contents(&immutable_host_block, &disk_blocks[0], 42)?;
1252
1253
1254
1255
1256

        Ok(())
    }

    #[tokio::test]
Ryan Olson's avatar
Ryan Olson committed
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
    #[rstest]
    #[case(LayoutType::FullyContiguous)]
    #[case(LayoutType::LayerSeparate { outer_contiguous: true })]
    #[case(LayoutType::LayerSeparate { outer_contiguous: false })]
    async fn test_onboard_disk(#[case] layout_type: LayoutType) -> Result<()> {
        let (offload_manager, device_pool, _, disk_pool) = build_pools_with_layout(
            4,
            None,
            Some(4),
            None,
            layout_type,
            BlockRegistrationDuplicationSetting::Disabled,
        )?;
1270

1271
1272
        let device_pool = device_pool.as_ref().unwrap();
        let disk_pool = disk_pool.as_ref().unwrap();
1273
1274
1275
1276
1277
1278
1279
1280
1281

        let disk_block = completed_block(disk_pool, [0, 1, 2, 3]).await?;
        let immutable_disk_block = disk_pool
            .register_blocks(vec![disk_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

1282
1283
        populate_block(&immutable_disk_block, 42)?;

1284
        let device_block = offload_manager
Ryan Olson's avatar
Ryan Olson committed
1285
1286
            .onboard(vec![immutable_disk_block.clone()], None)
            .await??;
1287

1288
1289
        check_block_contents(&immutable_disk_block, &device_block[0], 42)?;

1290
1291
        assert_eq!(device_block.len(), 1);
        assert_eq!(
Ryan Olson's avatar
Ryan Olson committed
1292
1293
            device_block[0].sequence_hash(),
            immutable_disk_block.sequence_hash()
1294
1295
1296
        );
        assert_eq!(
            device_pool
Ryan Olson's avatar
Ryan Olson committed
1297
                .match_sequence_hashes(vec![immutable_disk_block.sequence_hash()].as_slice())
1298
1299
1300
1301
1302
1303
1304
1305
1306
                .await?
                .len(),
            1
        );

        Ok(())
    }

    #[tokio::test]
Ryan Olson's avatar
Ryan Olson committed
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
    #[rstest]
    #[case(LayoutType::FullyContiguous)]
    #[case(LayoutType::LayerSeparate { outer_contiguous: true })]
    #[case(LayoutType::LayerSeparate { outer_contiguous: false })]
    async fn test_bulk_transfer_disk(#[case] layout_type: LayoutType) -> Result<()> {
        let (offload_manager, device_pool, host_pool, disk_pool) = build_pools_with_layout(
            8,
            Some(8),
            Some(8),
            None,
            layout_type,
            BlockRegistrationDuplicationSetting::Disabled,
        )?;
1320

1321
1322
1323
        let disk_pool = disk_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();
        let device_pool = device_pool.as_ref().unwrap();
1324
1325
1326
1327
1328

        let mut host_blocks = Vec::new();

        for i in 0..8 {
            let block = completed_block(host_pool, [i; 4]).await?;
1329
            populate_block(&block, i as u8)?;
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
            host_blocks.push(block);
        }

        let immutable_host_blocks = host_pool.register_blocks(host_blocks).await?;

        for block in &immutable_host_blocks {
            offload_manager.offload(block, 0).await?;
        }

        tokio::time::sleep(std::time::Duration::from_millis(500)).await;

        let mut disk_blocks = Vec::new();

1343
        for (i, host_block) in immutable_host_blocks.iter().enumerate() {
1344
            let blocks = disk_pool
Ryan Olson's avatar
Ryan Olson committed
1345
                .match_sequence_hashes(vec![host_block.sequence_hash()].as_slice())
1346
1347
                .await?;
            assert_eq!(blocks.len(), 1);
1348
            check_block_contents(host_block, &blocks[0], i as u8)?;
1349
1350
1351
            disk_blocks.push(blocks[0].clone());
        }

Ryan Olson's avatar
Ryan Olson committed
1352
        let device_blocks = offload_manager.onboard(disk_blocks.clone(), None).await??;
1353
1354
        assert_eq!(device_blocks.len(), disk_blocks.len());

1355
        for (i, disk_block) in disk_blocks.iter().enumerate() {
1356
            let blocks = device_pool
Ryan Olson's avatar
Ryan Olson committed
1357
                .match_sequence_hashes(vec![disk_block.sequence_hash()].as_slice())
1358
1359
                .await?;
            assert_eq!(blocks.len(), 1);
1360
            check_block_contents(disk_block, &blocks[0], i as u8)?;
1361
1362
1363
1364
        }

        Ok(())
    }
1365
1366
1367
1368
1369
1370
1371

    #[tokio::test]
    async fn test_transfer_batcher() -> Result<()> {
        let (offload_manager, device_pool, _, disk_pool) = build_pools(
            2 * MAX_TRANSFER_BATCH_SIZE + 1,
            None,
            Some(2 * MAX_TRANSFER_BATCH_SIZE + 1),
1372
            None,
1373
1374
1375
1376
1377
1378
1379
1380
        )?;

        let device_pool = device_pool.as_ref().unwrap();
        let disk_pool = disk_pool.as_ref().unwrap();

        let mut disk_blocks = Vec::new();

        for i in 0..2 * MAX_TRANSFER_BATCH_SIZE + 1 {
1381
1382
1383
            let disk_block = completed_block(disk_pool, [i as u32; 4]).await?;
            populate_block(&disk_block, i as u8)?;
            disk_blocks.push(disk_block);
1384
1385
1386
1387
1388
        }

        let immutable_disk_blocks = disk_pool.register_blocks(disk_blocks).await?;

        let device_blocks = offload_manager
Ryan Olson's avatar
Ryan Olson committed
1389
1390
            .onboard(immutable_disk_blocks.clone(), None)
            .await??;
1391
1392
        assert_eq!(device_blocks.len(), 2 * MAX_TRANSFER_BATCH_SIZE + 1);

1393
        for (i, device_block) in device_blocks.iter().enumerate() {
1394
            let blocks = device_pool
Ryan Olson's avatar
Ryan Olson committed
1395
                .match_sequence_hashes(vec![device_block.sequence_hash()].as_slice())
1396
                .await?;
1397
            check_block_contents(device_block, &blocks[0], i as u8)?;
1398
1399
1400
1401
1402
            assert_eq!(blocks.len(), 1);
        }

        Ok(())
    }
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418

    #[tokio::test]
    async fn test_onboard_unsupported_block_type() -> Result<()> {
        let (offload_manager, device_pool, _, _) = build_pools(1, None, None, None)?;

        let device_pool = device_pool.as_ref().unwrap();

        let block = completed_block(device_pool, [0; 4]).await?;

        let registered_block = device_pool
            .register_blocks(vec![block])
            .await?
            .into_iter()
            .next()
            .unwrap();

Ryan Olson's avatar
Ryan Olson committed
1419
1420
1421
        let onboarded_blocks = offload_manager
            .onboard(vec![registered_block], None)
            .await?;
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
        assert!(matches!(
            onboarded_blocks,
            Err(BlockPoolError::BlockError(BlockError::Other(_)))
        ));

        Ok(())
    }

    #[tokio::test]
    async fn test_offload_transfer_metadata() -> Result<()> {
        let (offload_manager, device_pool, host_pool, _) = build_pools(4, Some(4), None, None)?;

        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();

        let mut device_block = completed_block(device_pool, [0; 4]).await?;

        populate_block(&device_block, 42)?;

        let new_metadata = device_block.metadata().update_priority(1);
        device_block.update_metadata(new_metadata);

        let immutable_device_block = device_pool
            .register_blocks(vec![device_block])
            .await?
            .into_iter()
            .next()
            .unwrap();
        offload_manager.offload(&immutable_device_block, 0).await?;

        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        let host_blocks = host_pool
Ryan Olson's avatar
Ryan Olson committed
1455
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
            .await?;
        assert_eq!(host_blocks.len(), 1);
        check_block_contents(&immutable_device_block, &host_blocks[0], 42)?;
        assert_eq!(host_blocks[0].metadata().priority(), 1);

        Ok(())
    }

    #[tokio::test]
    async fn test_onboard_duplicate() -> Result<()> {
        let (offload_manager, device_pool, host_pool, _) = build_pools(4, Some(4), None, None)?;

        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();

        let device_block = completed_block(device_pool, [0; 4]).await?;

        let immutable_device_block = device_pool
            .register_blocks(vec![device_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

        populate_block(&immutable_device_block, 42)?;

        offload_manager.offload(&immutable_device_block, 0).await?;

        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        let host_blocks = host_pool
Ryan Olson's avatar
Ryan Olson committed
1487
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
1488
1489
1490
1491
            .await?;
        assert_eq!(host_blocks.len(), 1);

        let onboarded_blocks = offload_manager
Ryan Olson's avatar
Ryan Olson committed
1492
1493
            .onboard(vec![host_blocks[0].clone()], None)
            .await??;
1494
1495
1496
1497
1498
1499
        assert_eq!(onboarded_blocks.len(), 1);
        check_block_contents(&host_blocks[0], &onboarded_blocks[0], 42)?;

        // This should be the same block that we put on the device.
        // The block that was copied should be discarded by the block pool.
        assert_eq!(
Ryan Olson's avatar
Ryan Olson committed
1500
1501
            onboarded_blocks[0].block_id(),
            immutable_device_block.block_id()
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
        );

        Ok(())
    }

    #[tokio::test]
    async fn test_transfer_big_blocks() -> Result<()> {
        // Try a block size of 32 MB.
        let inner_dim = 2_usize.pow(20) * 32 / NUM_LAYERS / BLOCK_SIZE;
        let (offload_manager, device_pool, host_pool, disk_pool) =
            build_pools(2, Some(2), Some(2), Some(inner_dim))?;

        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();
        let disk_pool = disk_pool.as_ref().unwrap();

        let device_block = completed_block(device_pool, [0; 4]).await?;

        populate_block(&device_block, 42)?;

        let immutable_device_block = device_pool
            .register_blocks(vec![device_block])
            .await?
            .into_iter()
            .next()
            .unwrap();

        // Offload to host.
        offload_manager.offload(&immutable_device_block, 0).await?;

        // Wait for the offload to be processed.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        let host_blocks = host_pool
Ryan Olson's avatar
Ryan Olson committed
1536
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
            .await?;
        assert_eq!(host_blocks.len(), 1);
        check_block_contents(&immutable_device_block, &host_blocks[0], 42)?;

        // Offload to disk
        offload_manager.offload(&host_blocks[0], 0).await?;

        // Wait for the offload to be processed.
        tokio::time::sleep(std::time::Duration::from_millis(500)).await;

        let disk_blocks = disk_pool
Ryan Olson's avatar
Ryan Olson committed
1548
            .match_sequence_hashes(vec![immutable_device_block.sequence_hash()].as_slice())
1549
1550
1551
1552
1553
            .await?;
        assert_eq!(disk_blocks.len(), 1);
        check_block_contents(&host_blocks[0], &disk_blocks[0], 42)?;

        // Onboard to device.
Ryan Olson's avatar
Ryan Olson committed
1554
        let device_blocks = offload_manager.onboard(disk_blocks.clone(), None).await??;
1555
1556
1557
1558
1559
        assert_eq!(device_blocks.len(), 1);
        check_block_contents(&disk_blocks[0], &device_blocks[0], 42)?;

        Ok(())
    }
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596

    #[tokio::test]
    async fn test_offload_evict_order() -> Result<()> {
        let (offload_manager, device_pool, host_pool, _) = build_pools(4, Some(4), None, None)?;

        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();

        let tokens = vec![0_u32; BLOCK_SIZE * 4];
        let token_blocks = TokenBlockSequence::new(Tokens::from(tokens), 4, None);
        assert_eq!(token_blocks.blocks().len(), 4);

        let mut mutable_blocks = Vec::new();
        let mut sequence_hashes = Vec::new();
        for token_block in token_blocks.blocks() {
            let mut mutable_block = device_pool
                .allocate_blocks(1)
                .await?
                .into_iter()
                .next()
                .unwrap();
            mutable_block.apply_token_block(token_block.clone())?;
            sequence_hashes.push(mutable_block.sequence_hash()?);
            mutable_blocks.push(mutable_block);
        }

        let immutable_blocks = device_pool.register_blocks(mutable_blocks).await?;

        for block in &immutable_blocks {
            offload_manager.offload(block, 0).await?;
        }
        // Wait for offloads.
        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        // Allocate 2 blocks on the host.
        let _host_blocks = host_pool.allocate_blocks(2).await?;

Ryan Olson's avatar
Ryan Olson committed
1597
1598
        // The first two blocks should've been evicted.
        // The last two blocks should still be on the host.
1599
1600
1601
1602
1603
        assert_eq!(
            host_pool
                .match_sequence_hashes(sequence_hashes.as_slice())
                .await?
                .len(),
Ryan Olson's avatar
Ryan Olson committed
1604
            0
1605
1606
1607
1608
        );

        assert_eq!(
            host_pool
Ryan Olson's avatar
Ryan Olson committed
1609
                .match_sequence_hashes(&sequence_hashes[2..])
1610
1611
                .await?
                .len(),
Ryan Olson's avatar
Ryan Olson committed
1612
            2
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
        );

        Ok(())
    }

    #[tokio::test]
    async fn test_onboard_evict_order() -> Result<()> {
        let (offload_manager, device_pool, host_pool, _) = build_pools(4, Some(4), None, None)?;

        let device_pool = device_pool.as_ref().unwrap();
        let host_pool = host_pool.as_ref().unwrap();

        let tokens = vec![0_u32; BLOCK_SIZE * 4];
        let token_blocks = TokenBlockSequence::new(Tokens::from(tokens), 4, None);
        assert_eq!(token_blocks.blocks().len(), 4);

        let mut mutable_blocks = Vec::new();
        let mut sequence_hashes = Vec::new();
        for token_block in token_blocks.blocks() {
            let mut block = host_pool
                .allocate_blocks(1)
                .await?
                .into_iter()
                .next()
                .unwrap();
            block.apply_token_block(token_block.clone())?;

            sequence_hashes.push(block.sequence_hash()?);
            mutable_blocks.push(block);
        }

        let immutable_blocks = host_pool.register_blocks(mutable_blocks).await?;

Ryan Olson's avatar
Ryan Olson committed
1646
        let _ = offload_manager.onboard(immutable_blocks, None).await?;
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673

        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        let _device_blocks = device_pool.allocate_blocks(2).await?;

        assert_eq!(
            device_pool
                .match_sequence_hashes(sequence_hashes.as_slice())
                .await?
                .len(),
            2
        );

        tokio::time::sleep(std::time::Duration::from_millis(100)).await;

        let _device_blocks2 = device_pool.allocate_blocks(1).await?;

        assert_eq!(
            device_pool
                .match_sequence_hashes(sequence_hashes.as_slice())
                .await?
                .len(),
            1
        );

        Ok(())
    }
1674
}