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
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
163
164
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
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
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
654
655
656
657
658
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
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
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
814
815
816
817
818
819
820
821
822
823
824
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
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
use super::{
MemoryConfiguration, MemoryPoolOptions, MemoryUsage, PoolType,
memory_pool::{ExclusiveMemoryPool, MemoryPool, PersistentPool, SlicedPool},
};
use crate::{
config::{
CubeClRuntimeConfig, RuntimeConfig,
memory::{MemoryLogLevel, PersistentMemory},
},
logging::ServerLogger,
memory_management::{BytesFormat, memory_pool::Slice},
server::IoError,
storage::{ComputeStorage, StorageHandle},
};
use alloc::format;
use alloc::string::{String, ToString};
#[cfg(not(exclusive_memory_only))]
use alloc::vec;
use alloc::vec::Vec;
use cubecl_common::{backtrace::BackTrace, stub::Arc};
use cubecl_ir::MemoryDeviceProperties;
pub use super::memory_pool::{ManagedMemoryBinding, handle::*};
// These are 288 bytes vs 64 bytes. Adding boxing isn't really worth
// saving the 200 bytes.
#[allow(clippy::large_enum_variant)]
enum DynamicPool {
Sliced(SlicedPool),
Exclusive(ExclusiveMemoryPool),
}
impl MemoryPool for DynamicPool {
fn accept(&self, size: u64) -> bool {
match self {
DynamicPool::Sliced(pool) => pool.accept(size),
DynamicPool::Exclusive(pool) => pool.accept(size),
}
}
fn find(&self, binding: &ManagedMemoryBinding) -> Result<&Slice, IoError> {
match self {
DynamicPool::Sliced(m) => m.find(binding),
DynamicPool::Exclusive(m) => m.find(binding),
}
}
#[cfg_attr(feature = "tracing", tracing::instrument(level = "trace", skip(self)))]
fn try_reserve(&mut self, size: u64) -> Option<ManagedMemoryHandle> {
match self {
DynamicPool::Sliced(m) => m.try_reserve(size),
DynamicPool::Exclusive(m) => m.try_reserve(size),
}
}
#[cfg_attr(
feature = "tracing",
tracing::instrument(level = "trace", skip(self, storage))
)]
fn alloc<Storage: ComputeStorage>(
&mut self,
storage: &mut Storage,
size: u64,
) -> Result<ManagedMemoryHandle, IoError> {
match self {
DynamicPool::Sliced(m) => m.alloc(storage, size),
DynamicPool::Exclusive(m) => m.alloc(storage, size),
}
}
fn get_memory_usage(&self) -> MemoryUsage {
match self {
DynamicPool::Sliced(m) => m.get_memory_usage(),
DynamicPool::Exclusive(m) => m.get_memory_usage(),
}
}
fn cleanup<Storage: ComputeStorage>(
&mut self,
storage: &mut Storage,
alloc_nr: u64,
explicit: bool,
) {
match self {
DynamicPool::Sliced(m) => m.cleanup(storage, alloc_nr, explicit),
DynamicPool::Exclusive(m) => m.cleanup(storage, alloc_nr, explicit),
};
storage.flush();
}
fn bind(
&mut self,
reserved: ManagedMemoryHandle,
assigned: ManagedMemoryHandle,
cursor: u64,
) -> Result<(), IoError> {
match self {
DynamicPool::Sliced(m) => m.bind(reserved, assigned, cursor),
DynamicPool::Exclusive(m) => m.bind(reserved, assigned, cursor),
}
}
}
#[derive(Default, Clone, Copy, Debug)]
/// The mode of allocation used.
pub enum MemoryAllocationMode {
/// Use the automatic memory management strategy for allocation.
#[default]
Auto,
/// Use a persistent memory management strategy, meaning that all allocations are for data that is
/// likely never going to be freed.
Persistent,
}
/// What the capture arena is holding, and how it got there.
///
/// `misses` is the load-bearing one: a miss taken while a stream capture is
/// open is a driver allocation recorded as a graph MEMORY node, which is both a
/// per-replay cost and an address the next capture will not reproduce.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub struct ArenaStats {
/// How many times the arena has been opened.
pub generation: u64,
/// Reservations served from the arena since the last counter reset.
pub served: u64,
/// Of those, the ones that needed a fresh allocation from the storage.
pub misses: u64,
/// Slices the arena owns.
pub slices: u64,
/// Bytes the arena keeps reserved, whether in use or free.
pub bytes_reserved: u64,
/// Bytes of that currently held by a live handle.
pub bytes_in_use: u64,
}
/// Whether [`MemoryManagement::reserve`] times itself. Read once: a `getenv`
/// per reservation would be a cost on the path this exists to measure.
fn time_reserve() -> bool {
static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
*ON.get_or_init(|| std::env::var("CUBECL_TIME_RESERVE").as_deref() == Ok("1"))
}
/// Reserves and keeps track of chunks of memory in the storage, and slices upon these chunks.
pub struct MemoryManagement<Storage> {
name: String,
persistent: PersistentPool,
/// The capture-scoped arena. See [`MemoryManagement::arena_begin`].
arena: PersistentPool,
/// The `pool` byte an arena slice's location carries, so `find` and `bind`
/// can route back here. One past the persistent pool's.
arena_pool_pos: u8,
/// While true, EVERY `reserve` is served from `arena` and none from the
/// ordinary pools.
arena_active: bool,
/// Bumped on every [`MemoryManagement::arena_begin`]. Reporting only: what
/// decides whether two graphs may share this arena is `arena_sizes`, not
/// how many times it was opened.
arena_generation: u64,
/// Requests the arena served, and the subset it could not satisfy from an
/// already-owned slice. A miss inside a capture is exactly a graph memory
/// node, so this is the number the warm pass exists to drive to zero.
arena_served: u64,
arena_misses: u64,
/// Every size the arena has served since it was last opened, in order.
///
/// This is the region's IDENTITY as far as the arena is concerned. Two
/// captures that share an arena share its slices -- that is the point, it
/// is what makes their addresses comparable -- and sharing scratch is safe
/// exactly when both are captures of the same region, replayed serially.
/// The same region issues the same request sequence; a different one does
/// not, and a hash over the window is what lets a replay say so.
arena_sizes: Vec<u64>,
pools: Vec<DynamicPool>,
storage: Storage,
alloc_reserve_count: u64,
/// Wall time spent inside [`MemoryManagement::reserve`], in nanoseconds,
/// and how many calls it covers.
///
/// Off unless `CUBECL_TIME_RESERVE=1`, because it costs two clock reads a
/// reservation and a decode step makes about one per kernel launch. It
/// exists because "1869 allocation requests a step" is a COUNT, and a
/// count is not a time: at 0.5 us a request that is 0.9 ms of a 37 ms
/// enqueue and marginal, at 5 us it is 9.3 ms and it is the largest item
/// after the launches themselves. Nobody had any business asserting which
/// until it was timed.
///
/// FRAMING: this is the RESERVE half only. Returning a slice to the pool
/// happens on handle drop, elsewhere, and is not in this number.
reserve_nanos: u64,
reserve_calls: u64,
mode: MemoryAllocationMode,
config: PersistentMemory,
logger: Arc<ServerLogger>,
/// Externally-registered storage handles (e.g. GPU buffers aliasing mmap'd
/// host memory), keyed by memory-handle id. Resolved in `get_storage`,
/// bypassing the allocation pools entirely — never reserved, reused, or
/// reclaimed by the pool machinery. See [`Self::register_external`].
external: hashbrown::HashMap<usize, StorageHandle>,
}
fn generate_bucket_sizes(
start_size: u64,
end_size: u64,
max_buckets: usize,
alignment: u64,
) -> Vec<u64> {
let mut buckets = Vec::with_capacity(max_buckets);
let log_min = (start_size as f64).ln();
let log_max = (end_size as f64).ln();
let log_range = log_max - log_min;
// Pure exponential performed best, but let's try slightly denser in lower-mid range
for i in 0..max_buckets {
let p = i as f64 / (max_buckets - 1) as f64;
// Slight bias toward lower-mid range with less aggressive curve than sigmoid
let log_size = log_min + log_range * p;
let size = log_size.exp() as u64;
let aligned_size = size.next_multiple_of(alignment);
buckets.push(aligned_size);
}
buckets.dedup();
buckets
}
const DEALLOC_SCALE_MB: u64 = 1024 * 1024 * 1024;
const BASE_DEALLOC_PERIOD: u64 = 5000;
/// The options for creating a new [`MemoryManagement`] instance.
#[derive(Debug)]
pub struct MemoryManagementOptions {
/// The name of the memory management.
name: String,
/// The [`MemoryAllocationOption`] used by this instance.
memory: MemoryAllocationOption,
}
impl MemoryManagementOptions {
/// Creates a new [`MemoryManagementOptions`].
pub fn new<S: Into<String>>(name: S) -> Self {
Self {
name: name.into(),
memory: MemoryAllocationOption::FromConfig,
}
}
/// Forces the [`MemoryAllocationMode`] during execution to always be the provided one.
pub fn mode(mut self, mode: MemoryAllocationMode) -> Self {
self.memory = MemoryAllocationOption::Provided(mode);
self
}
}
#[derive(Default, Debug)]
/// Determines which [`MemoryAllocationMode`] is used during allocations.
enum MemoryAllocationOption {
#[default]
/// Uses the [`GlobalConfig`] to determine the mode of allocation.
FromConfig,
/// Use the provided [`MemoryAllocationMode`].
Provided(MemoryAllocationMode),
}
impl<Storage: ComputeStorage> MemoryManagement<Storage> {
/// Creates the options from device limits.
pub fn from_configuration(
storage: Storage,
properties: &MemoryDeviceProperties,
config: MemoryConfiguration,
logger: Arc<ServerLogger>,
options: MemoryManagementOptions,
) -> Self {
let pool_options = match config {
#[cfg(not(exclusive_memory_only))]
MemoryConfiguration::SubSlices => {
// Round chunk size to be aligned.
let memory_alignment = properties.alignment;
let max_page = properties.max_page_size;
let mut pools = Vec::new();
const MB: u64 = 1024 * 1024;
// Add in a pool for allocations that are smaller than the min alignment,
// as they can't use offsets at all (on wgpu at least).
pools.push(MemoryPoolOptions {
pool_type: PoolType::ExclusivePages { max_alloc_size: 0 },
dealloc_period: None,
});
let mut current = max_page;
let mut max_sizes = vec![];
let mut page_sizes = vec![];
let mut base = pools.len() as u32;
while current >= 32 * MB {
current /= 4;
// Make sure every pool has an aligned size.
current = current.next_multiple_of(memory_alignment);
max_sizes.push(current / 2u64.pow(base));
page_sizes.push(current);
base += 1;
}
max_sizes.reverse();
page_sizes.reverse();
for i in 0..max_sizes.len() {
let max = max_sizes[i];
let page_size = page_sizes[i];
pools.push(MemoryPoolOptions {
// Creating max slices lower than the chunk size reduces fragmentation.
pool_type: PoolType::SlicedPages {
page_size,
max_slice_size: max,
},
dealloc_period: None,
});
}
// Add pools from big to small.
pools.push(MemoryPoolOptions {
pool_type: PoolType::SlicedPages {
page_size: max_page / memory_alignment * memory_alignment,
max_slice_size: max_page / memory_alignment * memory_alignment,
},
dealloc_period: None,
});
pools
}
MemoryConfiguration::ExclusivePages => {
// Add all bin sizes. Nb: because of alignment some buckets
// end up as the same size, so only want unique ones,
// but also keep the order, so a BTree will do.
const MIN_BUCKET_SIZE: u64 = 1024 * 32;
const NUM_POOLS: usize = 24;
let sizes = generate_bucket_sizes(
MIN_BUCKET_SIZE,
properties.max_page_size,
NUM_POOLS,
properties.alignment,
);
sizes
.iter()
.map(|&size| {
let dealloc_period = (BASE_DEALLOC_PERIOD as f64
* (1.0 + size as f64 / (DEALLOC_SCALE_MB as f64)).round())
as u64;
MemoryPoolOptions {
pool_type: PoolType::ExclusivePages {
max_alloc_size: size,
},
dealloc_period: Some(dealloc_period),
}
})
.collect()
}
MemoryConfiguration::Custom { pool_options } => pool_options,
};
logger.log_memory(
|level| !matches!(level, MemoryLogLevel::Disabled),
|| {
let mut msg = String::new();
for pool in pool_options.iter() {
msg += &format!("[{}] Using memory pool: \n {pool:?}\n", options.name);
}
msg
},
);
let pools: Vec<_> = pool_options
.iter()
.enumerate()
.map(|(pool_pos, options)| {
let pool_pos = pool_pos as u8;
match options.pool_type {
PoolType::SlicedPages {
page_size,
max_slice_size,
} => DynamicPool::Sliced(SlicedPool::new(
page_size,
max_slice_size,
properties.alignment,
pool_pos,
)),
PoolType::ExclusivePages { max_alloc_size } => {
DynamicPool::Exclusive(ExclusiveMemoryPool::new(
max_alloc_size,
properties.alignment,
options.dealloc_period.unwrap_or(u64::MAX),
pool_pos,
))
}
}
})
.collect();
let config = CubeClRuntimeConfig::get().memory.persistent_memory.clone();
let mode = match options.memory {
MemoryAllocationOption::Provided(mode) => mode,
MemoryAllocationOption::FromConfig => match config {
PersistentMemory::Enabled => MemoryAllocationMode::Auto,
PersistentMemory::Disabled => MemoryAllocationMode::Auto,
PersistentMemory::Enforced => MemoryAllocationMode::Persistent,
},
};
Self {
name: options.name,
persistent: PersistentPool::new(
properties.max_page_size,
properties.alignment,
pools.len() as u8,
),
// The arena accepts any size the storage will accept: a captured
// region does not get to choose which of its buffers are small.
arena: PersistentPool::new(u64::MAX, properties.alignment, pools.len() as u8 + 1),
arena_pool_pos: pools.len() as u8 + 1,
arena_active: false,
arena_generation: 0,
arena_served: 0,
arena_misses: 0,
arena_sizes: Vec::new(),
pools,
storage,
alloc_reserve_count: 0,
reserve_nanos: 0,
reserve_calls: 0,
mode,
config,
logger,
external: hashbrown::HashMap::new(),
}
}
/// Change the mode of allocation.
pub fn mode(&mut self, mode: MemoryAllocationMode) {
// We override the mode based on the cubecl config.
let mode = match self.config {
PersistentMemory::Enabled => mode,
PersistentMemory::Disabled | PersistentMemory::Enforced => return,
};
self.logger.log_memory(
|level| !matches!(level, MemoryLogLevel::Disabled),
|| {
format!(
"[{}] Setting memory allocation mode: from {:?} => {mode:?}",
self.name, self.mode
)
},
);
self.mode = mode;
}
/// Route every subsequent `reserve` into the capture arena, and return the
/// generation this opening is.
///
/// The arena exists because a captured graph records ADDRESSES. Three
/// things have to be true at once, and no ordinary pool gives all three:
///
/// 1. **Reuse inside the region.** A buffer born inside the captured region
/// and dead inside it must hand its slice back, so the region's peak
/// footprint is its live set and not its allocation count. Without this
/// every intra-region allocation becomes a driver allocation made while
/// the stream is capturing, i.e. a graph MEMORY node.
/// 2. **Whole-arena reservation.** Once the region is captured, every slice
/// the arena ever handed out is baked into a node, so NOTHING outside a
/// capture may be given one for as long as the graph can be replayed.
/// The arena is only ever consulted while it is open, and it never
/// returns pages to the storage, so this holds for the whole arena and
/// not just for the slices that happened to be live at capture end.
/// 3. **A deterministic base.** Reopening the arena resets no slice and
/// frees nothing; slices are matched by exact effective size in a fixed
/// order. A periodic region therefore issues an identical request
/// sequence against an identical free set and gets identical addresses,
/// which is what makes two captures of the same region comparable and a
/// single graph legitimate for every step.
///
/// Note what this is NOT: it is not "hold the buffers a capture touched".
/// Holding gets reservation right and reuse wrong -- it is precisely what
/// stops (1) -- and holding only the buffers that were live when the region
/// opened gets reuse right and reservation wrong for everything born inside
/// it. The two properties have to be separated onto different memory, which
/// is what the arena is.
///
/// Buffers that were live when the arena opened are NOT in it. They belong
/// to the ordinary pools, and if one dies inside the region its slice goes
/// back to a pool the region is not allocating from -- so it cannot be
/// handed to a later node of the same region. That is the whole of the
/// intra-region aliasing bug, closed structurally rather than by holding.
/// The caller still owes such a buffer a hold for the graph's lifetime, to
/// stop the pool handing it out AFTER the capture closes.
///
/// TWO GRAPHS MAY SHARE ONE ARENA, and under (3) they positively should:
/// that is what makes their recorded addresses comparable, and it is what
/// lets a graph captured on one step stand in for the next. They then share
/// scratch, which is safe exactly when they are captures of the SAME region
/// and are replayed serially. Serial is a property of the one stream they
/// replay on; same-region is what [`Self::arena_signature`] is for.
pub fn arena_begin(&mut self) -> u64 {
self.arena_generation += 1;
self.arena_active = true;
self.arena_sizes.clear();
self.logger.log_memory(
|level| !matches!(level, MemoryLogLevel::Disabled),
|| {
format!(
"[{}] Capture arena open, generation {}",
self.name, self.arena_generation
)
},
);
self.arena_generation
}
/// Stop routing `reserve` into the arena. The arena keeps everything it
/// owns: closing it is the end of allocation from it, never the end of its
/// reservation.
pub fn arena_end(&mut self) {
self.arena_active = false;
}
/// Whether `reserve` is currently being served from the arena.
pub fn arena_active(&self) -> bool {
self.arena_active
}
/// What the arena is holding, and how it got there.
pub fn arena_stats(&self) -> ArenaStats {
let usage = self.arena.get_memory_usage();
ArenaStats {
generation: self.arena_generation,
served: self.arena_served,
misses: self.arena_misses,
slices: self.arena.slice_count() as u64,
bytes_reserved: usage.bytes_reserved,
bytes_in_use: usage.bytes_in_use,
}
}
/// Zero the served/miss counters, so a capture pass can be counted apart
/// from the warm pass that preceded it. Does not free or reset anything.
pub fn arena_reset_counters(&mut self) {
self.arena_served = 0;
self.arena_misses = 0;
}
/// Give the arena's free pages back to the storage.
///
/// Only legal once no captured graph built against it can still be
/// replayed -- every such graph holds pointers into these pages. Slices
/// still held by a live handle are kept.
pub fn arena_release(&mut self) {
assert!(
!self.arena_active,
"[{}] arena_release while the arena is open",
self.name
);
self.arena
.cleanup(&mut self.storage, self.alloc_reserve_count, true);
}
/// Whether this binding's memory came from the arena.
pub fn arena_owns(&self, binding: &ManagedMemoryBinding) -> bool {
binding.descriptor().location().pool == self.arena_pool_pos
}
/// How many requests the open arena window has served so far. A caller
/// brackets a region with two of these to name the window it owns.
pub fn arena_mark(&self) -> usize {
self.arena_sizes.len()
}
/// A hash of every size served since `from`, in order: the signature of the
/// region that allocated them. See [`Self::arena_sizes`].
pub fn arena_signature(&self, from: usize) -> u64 {
let mut h: u64 = 0xcbf2_9ce4_8422_2325;
for size in &self.arena_sizes[from.min(self.arena_sizes.len())..] {
h ^= *size;
h = h.wrapping_mul(0x1000_0000_01b3);
}
h
}
fn arena_reserve(&mut self, size: u64) -> Result<ManagedMemoryHandle, IoError> {
self.arena_served += 1;
self.arena_sizes.push(size);
if let Some(handle) = self.arena.try_reserve(size) {
return Ok(handle);
}
// A miss is a real driver allocation. Inside a capture that is a graph
// memory node; the warm pass exists so this branch is not taken then.
self.arena_misses += 1;
self.arena.alloc(&mut self.storage, size)
}
/// Cleanup allocations in pools that are deemed unnecessary.
pub fn cleanup(&mut self, explicit: bool) {
self.logger.log_memory(
|level| !matches!(level, MemoryLogLevel::Disabled) && explicit,
|| "Manual memory cleanup ...".to_string(),
);
self.persistent
.cleanup(&mut self.storage, self.alloc_reserve_count, explicit);
for pool in self.pools.iter_mut() {
pool.cleanup(&mut self.storage, self.alloc_reserve_count, explicit);
}
}
/// Returns the storage from the specified binding
pub fn get_cursor(&self, binding: ManagedMemoryBinding) -> Result<u64, IoError> {
// External storage is registered rather than reserved: it has no
// `Slice`, so it has no cursor to read. Returning 0 is not a fallback,
// it is the right answer -- the cursor exists to order a read after the
// stream operation that last WROTE the memory, and external storage is
// pinned `can_mut() == false` precisely so that nothing ever writes it.
// There is no producing operation to wait for.
//
// Only the CUDA backend reaches this: its multi-stream scheduler calls
// `handle_cursor` on every binding, while the wgpu backend the seam was
// first written for has no cursors at all. That is why an external
// handle used to die here with "Memory page 0 doesn't exist".
if self.external.contains_key(&binding.descriptor().id.value) {
return Ok(0);
}
let slice = self.find(binding)?;
Ok(slice.cursor)
}
/// Returns the storage from the specified binding
fn find(&self, binding: ManagedMemoryBinding) -> Result<&Slice, IoError> {
let id = binding.descriptor();
if id.location().pool == self.arena_pool_pos {
return self.arena.find(&binding);
}
if id.location().pool >= self.pools.len() as u8 {
return self.persistent.find(&binding);
}
let pool =
self.pools
.get(id.location().pool as usize)
.ok_or_else(|| IoError::NotFound {
backtrace: BackTrace::capture(),
reason: format!("Pool {} doesn't exist", id.location().pool).into(),
})?;
let slice = pool.find(&binding)?;
assert_eq!(slice.handle.descriptor(), binding.descriptor());
Ok(slice)
}
/// Returns the storage from the specified binding
pub fn get_storage(&mut self, binding: ManagedMemoryBinding) -> Result<StorageHandle, IoError> {
if let Some(storage) = self.external.get(&binding.descriptor().id.value) {
return Ok(storage.clone());
}
let slice = self.find(binding)?;
Ok(slice.storage.clone())
}
/// Register an externally-created [`StorageHandle`] (e.g. a GPU buffer that
/// aliases mmap'd host memory) as a tensor-addressable memory handle,
/// BYPASSING the allocation pools. The returned handle resolves to `storage`
/// via [`Self::get_storage`]/[`Self::get_resource`] but is never reserved,
/// reused, sub-sliced, or reclaimed — the caller owns the underlying storage
/// lifetime. This is the zero-copy seam for weights-as-tribles: an mmap'd f16
/// pile page becomes a GPU tensor with no copy and no allocation.
///
/// The returned handle (and every clone of it) permanently reports
/// `can_mut() == false`: external storage aliases READ-ONLY host memory,
/// so it must never be picked as an in-place kernel destination. Burn's
/// elementwise kernels reuse an "owned" input handle as the output buffer
/// (`can_mut()` = a handle-count heuristic, `strong_count <= 2`); on a
/// read-only mmap the GPU's in-place write is silently dropped by the
/// pager and the op returns its INPUT bytes — e.g. voxtral's folded-ear
/// load saw `q.mul_scalar(1/sqrt(d))` come back unscaled (2026-07-12,
/// caught by the aliased-vs-materialized A/B being token-divergent).
/// Pinning two extra handle clones for the registration's life (external
/// entries are never removed anyway) pushes every user-facing clone over
/// the heuristic's threshold, forcing all consumers to allocate outputs.
pub fn register_external(&mut self, storage: StorageHandle) -> ManagedMemoryHandle {
let handle = ManagedMemoryHandle::new();
core::mem::forget(handle.clone());
core::mem::forget(handle.clone());
self.external.insert(handle.descriptor().id.value, storage);
handle
}
/// Returns the resource from the storage at the specified handle
pub fn get_resource(
&mut self,
binding: ManagedMemoryBinding,
offset_start: Option<u64>,
offset_end: Option<u64>,
) -> Result<Storage::Resource, IoError> {
let handle = self.get_storage(binding)?;
let handle = match offset_start {
Some(offset) => handle.offset_start(offset),
None => handle,
};
let handle = match offset_end {
Some(offset) => handle.offset_end(offset),
None => handle,
};
Ok(self.storage().get(&handle))
}
/// Finds a spot in memory for a resource with the given size in bytes, and returns a handle to it
#[cfg_attr(feature = "tracing", tracing::instrument(level = "trace", skip(self)))]
pub fn reserve(&mut self, size: u64) -> Result<ManagedMemoryHandle, IoError> {
if time_reserve() {
let t = web_time::Instant::now();
let out = self.reserve_inner(size);
self.reserve_nanos += t.elapsed().as_nanos() as u64;
self.reserve_calls += 1;
return out;
}
self.reserve_inner(size)
}
/// What a reservation costs, as `(calls, nanoseconds)` since the last
/// reset. Zero unless `CUBECL_TIME_RESERVE=1`.
pub fn reserve_timing(&self) -> (u64, u64) {
(self.reserve_calls, self.reserve_nanos)
}
/// Zero the reservation timer, so one pass can be counted apart from the
/// startup that preceded it.
pub fn reserve_timing_reset(&mut self) {
self.reserve_calls = 0;
self.reserve_nanos = 0;
}
fn reserve_inner(&mut self, size: u64) -> Result<ManagedMemoryHandle, IoError> {
// If this happens every nanosecond, counts overflows after 585 years, so not worth thinking too
// hard about overflow here.
self.alloc_reserve_count += 1;
// The arena is not one pool among the others: while it is open it is
// the ONLY one, because an address a captured node records has to come
// from memory nothing outside the capture can be given.
if self.arena_active {
return self.arena_reserve(size);
}
if let Some(val) = self.persistent.try_reserve(size) {
self.logger.log_memory(
|level| matches!(level, MemoryLogLevel::Full),
|| {
format!(
"[{}] Reserved memory {size} using persistent memory",
self.name
)
},
);
return Ok(val);
}
if matches!(self.mode, MemoryAllocationMode::Persistent) || self.persistent.has_size(size) {
let allocated = self.persistent.alloc(&mut self.storage, size);
self.logger.log_memory(
|level| !matches!(level, MemoryLogLevel::Disabled),
|| {
format!(
"[{}] Allocated a new memory page using persistent memory, \n{}",
self.name, self,
)
},
);
return allocated;
}
self.logger.log_memory(
|level| matches!(level, MemoryLogLevel::Full),
|| {
format!(
"[{}] Reserved memory {} using dynamic pool",
self.name,
BytesFormat::new(size)
)
},
);
// Find first pool that fits this allocation
let pool = self
.pools
.iter_mut()
.find(|p| p.accept(size))
.ok_or(IoError::BufferTooBig {
size,
backtrace: BackTrace::capture(),
})?;
if let Some(slice) = pool.try_reserve(size) {
return Ok(slice);
}
let allocated = pool.alloc(&mut self.storage, size);
self.logger.log_memory(
|level| matches!(level, MemoryLogLevel::Full),
|| {
format!(
"[{}], Allocated a new memory page, current usage: \n{}",
self.name, self
)
},
);
allocated
}
/// Fetch the storage used by the memory manager.
///
/// # Notes
///
/// The storage should probably not be used for allocations since the handles won't be
/// compatible with the ones provided by the current trait. Prefer using the
/// [alloc](ComputeStorage::alloc) and [dealloc](ComputeStorage::dealloc) functions.
///
/// This is useful if you need to time the deallocations based on async computation, or to
/// change the mode of storage for different reasons.
pub fn storage(&mut self) -> &mut Storage {
&mut self.storage
}
/// Get the current memory usage.
pub fn memory_usage(&self) -> MemoryUsage {
let memory_usage = self.pools.iter().map(|x| x.get_memory_usage()).fold(
MemoryUsage {
number_allocs: 0,
bytes_in_use: 0,
bytes_padding: 0,
bytes_reserved: 0,
},
|m1, m2| m1.combine(m2),
);
memory_usage
.combine(self.persistent.get_memory_usage())
.combine(self.arena.get_memory_usage())
}
/// Print out a report of the current memory usage.
pub fn print_memory_usage(&self) {
#[cfg(feature = "std")]
log::info!("{}", self.memory_usage());
}
/// Binds the given [handle](HandleId) to a [`MemorySlot`].
pub fn bind(
&mut self,
reserved: ManagedMemoryHandle,
assigned: ManagedMemoryHandle,
cursor: u64,
) -> Result<(), IoError> {
let descriptor = reserved.descriptor();
if descriptor.location().init == 0 {
return Err(IoError::NotFound {
backtrace: BackTrace::capture(),
reason: "Reserved memory isn't initialized".into(),
});
}
let pool_index = descriptor.location().pool as usize;
if pool_index == self.arena_pool_pos as usize {
return self.arena.bind(reserved, assigned, cursor);
}
if pool_index >= self.pools.len() {
return self.persistent.bind(reserved, assigned, cursor);
}
self.pools
.get_mut(pool_index)
.map(|p| p.bind(reserved, assigned, cursor))
.ok_or_else(|| IoError::NotFound {
backtrace: BackTrace::capture(),
reason: format!("Memory pool {} doesn't exist", pool_index).into(),
})?
}
}
impl<Storage: ComputeStorage> core::fmt::Display for MemoryManagement<Storage> {
fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
f.write_str("\n# MemoryManagement\n\n")?;
f.write_fmt(format_args!(" - name: {:?}\n", self.name))?;
f.write_fmt(format_args!("\n## Persistent\n\n{}", self.persistent))?;
f.write_fmt(format_args!(
"\n## Capture arena (generation {}, {})\n\n{}",
self.arena_generation,
match self.arena_active {
true => "open",
false => "closed",
},
self.arena
))?;
f.write_str("\n## Dynamic\n\n")?;
for pool in self.pools.iter() {
match pool {
DynamicPool::Sliced(pool) => f.write_fmt(format_args!("{pool}\n"))?,
DynamicPool::Exclusive(pool) => f.write_fmt(format_args!("{pool}\n"))?,
}
}
let memory_usage = self.memory_usage();
f.write_fmt(format_args!("\n## Summary\n\n{memory_usage}"))?;
Ok(())
}
}
impl<Storage> core::fmt::Debug for MemoryManagement<Storage> {
fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
f.write_str(
alloc::format!(
"DynamicMemoryManagement {:?}",
core::any::type_name::<Storage>(),
)
.as_str(),
)
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::{memory_management::MemoryManagement, storage::BytesStorage};
use alloc::vec;
const DUMMY_MEM_PROPS: MemoryDeviceProperties = MemoryDeviceProperties {
max_page_size: 128 * 1024 * 1024,
alignment: 32,
};
fn options() -> MemoryManagementOptions {
MemoryManagementOptions {
name: "test".into(),
memory: MemoryAllocationOption::FromConfig,
}
}
// Test pools with slices.
#[test_log::test]
#[cfg(not(exclusive_memory_only))]
fn test_handle_mutability() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::SubSlices,
Arc::new(ServerLogger::default()),
options(),
);
let handle = memory_management.reserve(10).unwrap();
let other_ref = handle.clone();
assert!(!handle.can_mut(), "Handle can't be mut when multiple ref.");
drop(other_ref);
assert!(handle.can_mut(), "Handle should be mut when only one ref.");
}
// Test pools with slices.
#[test_log::test]
#[cfg(not(exclusive_memory_only))]
fn test_memory_usage() {
let max_page_size = 512;
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::ExclusivePages {
max_alloc_size: max_page_size,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let handle = memory_management.reserve(100);
let usage = memory_management.memory_usage();
assert_eq!(usage.bytes_in_use, 100);
assert!(usage.bytes_reserved >= 100 && usage.bytes_reserved <= max_page_size);
// Drop and re-alloc.
drop(handle);
let _handle = memory_management.reserve(100);
let usage_new = memory_management.memory_usage();
assert_eq!(usage, usage_new);
}
#[test_log::test]
fn alloc_two_chunks_on_one_page() {
let page_size = 2048;
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::SlicedPages {
page_size,
max_slice_size: page_size,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 512;
let _handle = memory_management.reserve(alloc_size);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
assert_eq!(usage.number_allocs, 2);
assert_eq!(usage.bytes_in_use, alloc_size * 2);
assert_eq!(usage.bytes_reserved, page_size);
}
#[test_log::test]
fn alloc_reuses_storage() {
// If no storage is re-used, this will allocate two pages.
let page_size = 512;
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::SlicedPages {
page_size,
max_slice_size: page_size,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 512;
let _handle = memory_management.reserve(alloc_size);
drop(_handle);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
assert_eq!(usage.number_allocs, 1);
assert_eq!(usage.bytes_in_use, alloc_size);
assert_eq!(usage.bytes_reserved, page_size);
}
#[test_log::test]
fn alloc_allocs_new_storage() {
let page_size = 1024;
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::SlicedPages {
page_size,
max_slice_size: page_size,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 768;
let _handle = memory_management.reserve(alloc_size);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
assert_eq!(usage.number_allocs, 2);
assert_eq!(usage.bytes_in_use, alloc_size * 2);
assert_eq!(usage.bytes_reserved, page_size * 2);
}
#[test_log::test]
fn alloc_respects_alignment_size() {
let page_size = 500;
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&MemoryDeviceProperties {
max_page_size: page_size,
alignment: 50,
},
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::SlicedPages {
page_size,
max_slice_size: page_size,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 40;
let _handle = memory_management.reserve(alloc_size);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
// Each slice should be aligned to 50 bytes, so 20 padding bytes.
assert_eq!(usage.bytes_padding, 10 * 2);
}
#[test_log::test]
fn allocs_on_correct_page() {
let sizes = [100, 200, 300, 400];
let pools = sizes
.iter()
.map(|size| MemoryPoolOptions {
pool_type: PoolType::SlicedPages {
page_size: *size,
max_slice_size: *size,
},
dealloc_period: None,
})
.collect();
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&MemoryDeviceProperties {
max_page_size: 128 * 1024 * 1024,
alignment: 10,
},
MemoryConfiguration::Custom {
pool_options: pools,
},
Arc::new(ServerLogger::default()),
options(),
);
// Allocate one thing on each page.
let alloc_sizes = [50, 150, 250, 350];
let _handles = alloc_sizes.map(|s| memory_management.reserve(s));
let usage = memory_management.memory_usage();
// Total memory should be size of all pages, and no more.
assert_eq!(usage.bytes_in_use, alloc_sizes.iter().sum::<u64>());
assert!(usage.bytes_reserved >= sizes.iter().sum::<u64>());
}
#[test_log::test]
#[cfg(not(exclusive_memory_only))]
fn allocate_deallocate_reallocate() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&MemoryDeviceProperties {
max_page_size: 128 * 1024 * 1024,
alignment: 32,
},
MemoryConfiguration::SubSlices,
Arc::new(ServerLogger::default()),
options(),
);
// Allocate a bunch
let handles: Vec<_> = (0..5)
.map(|i| memory_management.reserve(1000 * (i + 1)))
.collect();
let usage_before = memory_management.memory_usage();
// Deallocate
drop(handles);
// Reallocate
let _new_handles: Vec<_> = (0..5)
.map(|i| memory_management.reserve(1000 * (i + 1)))
.collect();
let usage_after = memory_management.memory_usage();
assert_eq!(usage_before.number_allocs, usage_after.number_allocs);
assert_eq!(usage_before.bytes_in_use, usage_after.bytes_in_use);
// Usage after can actually be _less_ because of defragging.
assert!(usage_before.bytes_reserved >= usage_after.bytes_reserved);
}
#[test_log::test]
#[cfg(not(exclusive_memory_only))]
fn test_fragmentation_resistance() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&MemoryDeviceProperties {
max_page_size: 128 * 1024 * 1024,
alignment: 32,
},
MemoryConfiguration::SubSlices,
Arc::new(ServerLogger::default()),
options(),
);
// Allocate a mix of small and large chunks
let sizes = [50, 1000, 100, 5000, 200, 10000, 300];
let handles: Vec<_> = sizes
.iter()
.map(|&size| memory_management.reserve(size).unwrap())
.collect();
let usage_before = memory_management.memory_usage();
// Deallocate every other allocation
for i in (0..handles.len()).step_by(2) {
drop(handles[i].clone());
}
// Reallocate similar sizes
for &size in &sizes[0..sizes.len() / 2] {
memory_management.reserve(size).unwrap();
}
let usage_after = memory_management.memory_usage();
// Check that we haven't increased our memory usage significantly
assert!(usage_after.bytes_reserved <= (usage_before.bytes_reserved as f64 * 1.1) as u64);
}
// Test pools without slices. More or less same as tests above.
#[test_log::test]
fn noslice_test_handle_mutability() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&(MemoryDeviceProperties {
max_page_size: 128 * 1024 * 1024,
alignment: 32,
}),
MemoryConfiguration::ExclusivePages,
Arc::new(ServerLogger::default()),
options(),
);
let handle = memory_management.reserve(10).unwrap();
let other_ref = handle.clone();
assert!(!handle.can_mut(), "Handle can't be mut when multiple ref.");
drop(other_ref);
assert!(handle.can_mut(), "Handle should be mut when only one ref.");
}
#[test_log::test]
fn noslice_alloc_two_chunk() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::ExclusivePages {
max_alloc_size: 1024,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 512;
let _handle = memory_management.reserve(alloc_size);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
assert_eq!(usage.number_allocs, 2);
assert_eq!(usage.bytes_in_use, alloc_size * 2);
assert!(usage.bytes_reserved >= alloc_size * 2);
}
#[test_log::test]
fn noslice_alloc_reuses_storage() {
// If no storage is re-used, this will allocate two pages.
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::ExclusivePages {
max_alloc_size: 1024,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 512;
let _handle = memory_management.reserve(alloc_size);
drop(_handle);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
assert_eq!(usage.number_allocs, 1);
assert_eq!(usage.bytes_in_use, alloc_size);
assert!(usage.bytes_reserved >= alloc_size);
}
#[test_log::test]
fn noslice_alloc_allocs_new_storage() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&DUMMY_MEM_PROPS,
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::ExclusivePages {
max_alloc_size: 1024,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 768;
let _handle = memory_management.reserve(alloc_size);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
assert_eq!(usage.number_allocs, 2);
assert_eq!(usage.bytes_in_use, alloc_size * 2);
assert!(usage.bytes_reserved >= alloc_size * 2);
}
#[test_log::test]
fn noslice_alloc_respects_alignment_size() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&MemoryDeviceProperties {
max_page_size: DUMMY_MEM_PROPS.max_page_size,
alignment: 50,
},
MemoryConfiguration::Custom {
pool_options: vec![MemoryPoolOptions {
pool_type: PoolType::ExclusivePages {
max_alloc_size: 50 * 20,
},
dealloc_period: None,
}],
},
Arc::new(ServerLogger::default()),
options(),
);
let alloc_size = 40;
let _handle = memory_management.reserve(alloc_size);
let _new_handle = memory_management.reserve(alloc_size);
let usage = memory_management.memory_usage();
// Each slice should be aligned to 60 bytes, so 20 padding bytes.
assert_eq!(usage.bytes_padding, 10 * 2);
}
#[test_log::test]
fn noslice_allocs_on_correct_page() {
let pools = [100, 200, 300, 400]
.iter()
.map(|&size| MemoryPoolOptions {
pool_type: PoolType::SlicedPages {
page_size: size,
max_slice_size: size,
},
dealloc_period: None,
})
.collect();
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&MemoryDeviceProperties {
max_page_size: DUMMY_MEM_PROPS.max_page_size,
alignment: 10,
},
MemoryConfiguration::Custom {
pool_options: pools,
},
Arc::new(ServerLogger::default()),
options(),
);
// Allocate one thing on each page.
let alloc_sizes = [50, 150, 250, 350];
let _handles = alloc_sizes.map(|s| memory_management.reserve(s));
let usage = memory_management.memory_usage();
// Total memory should be size of all pages, and no more.
assert_eq!(usage.bytes_in_use, alloc_sizes.iter().sum::<u64>());
}
#[test_log::test]
fn noslice_allocate_deallocate_reallocate() {
let mut memory_management = MemoryManagement::from_configuration(
BytesStorage::default(),
&MemoryDeviceProperties {
max_page_size: 128 * 1024 * 1024,
alignment: 32,
},
MemoryConfiguration::ExclusivePages,
Arc::new(ServerLogger::default()),
options(),
);
// Allocate a bunch
let handles: Vec<_> = (0..5)
.map(|i| memory_management.reserve(1000 * (i + 1)))
.collect();
let usage_before = memory_management.memory_usage();
// Deallocate
drop(handles);
// Reallocate
let _new_handles: Vec<_> = (0..5)
.map(|i| memory_management.reserve(1000 * (i + 1)))
.collect();
let usage_after = memory_management.memory_usage();
assert_eq!(usage_before.number_allocs, usage_after.number_allocs);
assert_eq!(usage_before.bytes_in_use, usage_after.bytes_in_use);
assert_eq!(usage_before.bytes_reserved, usage_after.bytes_reserved);
}
}