test_mp_dataloader.py 26.7 KB
Newer Older
1
import multiprocessing as mp
Rhett Ying's avatar
Rhett Ying committed
2
3
import os
import tempfile
4
import time
5
import unittest
Rhett Ying's avatar
Rhett Ying committed
6

7
import backend as F
Rhett Ying's avatar
Rhett Ying committed
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
import dgl
import numpy as np
import pytest
import torch as th
from dgl.data import CitationGraphDataset
from dgl.distributed import (
    DistDataLoader,
    DistGraph,
    DistGraphServer,
    load_partition,
    partition_graph,
)
from scipy import sparse as spsp
from utils import generate_ip_config, reset_envs

23
24

class NeighborSampler(object):
25
26
27
28
29
30
31
32
    def __init__(
        self,
        g,
        fanouts,
        sample_neighbors,
        use_graphbolt=False,
        return_eids=False,
    ):
33
34
35
        self.g = g
        self.fanouts = fanouts
        self.sample_neighbors = sample_neighbors
36
37
        self.use_graphbolt = use_graphbolt
        self.return_eids = return_eids
38
39
40

    def sample_blocks(self, seeds):
        import torch as th
Rhett Ying's avatar
Rhett Ying committed
41

42
        seeds = th.tensor(np.asarray(seeds), dtype=self.g.idtype)
43
44
45
46
        blocks = []
        for fanout in self.fanouts:
            # For each seed node, sample ``fanout`` neighbors.
            frontier = self.sample_neighbors(
47
                self.g, seeds, fanout, use_graphbolt=self.use_graphbolt
Rhett Ying's avatar
Rhett Ying committed
48
49
50
            )
            # Then we compact the frontier into a bipartite graph for
            # message passing.
51
52
53
            block = dgl.to_block(frontier, seeds)
            # Obtain the seed nodes for next layer.
            seeds = block.srcdata[dgl.NID]
54
55
56
            if frontier.num_edges() > 0:
                if not self.use_graphbolt or self.return_eids:
                    block.edata[dgl.EID] = frontier.edata[dgl.EID]
57
58
59
60
61

            blocks.insert(0, block)
        return blocks


Rhett Ying's avatar
Rhett Ying committed
62
63
64
65
66
67
def start_server(
    rank,
    ip_config,
    part_config,
    disable_shared_mem,
    num_clients,
68
    use_graphbolt=False,
Rhett Ying's avatar
Rhett Ying committed
69
70
71
72
73
74
75
76
77
78
):
    print("server: #clients=" + str(num_clients))
    g = DistGraphServer(
        rank,
        ip_config,
        1,
        num_clients,
        part_config,
        disable_shared_mem=disable_shared_mem,
        graph_format=["csc", "coo"],
79
        use_graphbolt=use_graphbolt,
Rhett Ying's avatar
Rhett Ying committed
80
    )
81
82
83
    g.start()


Rhett Ying's avatar
Rhett Ying committed
84
85
86
87
88
89
90
91
def start_dist_dataloader(
    rank,
    ip_config,
    part_config,
    num_server,
    drop_last,
    orig_nid,
    orig_eid,
92
93
    use_graphbolt=False,
    return_eids=False,
Rhett Ying's avatar
Rhett Ying committed
94
95
):
    dgl.distributed.initialize(ip_config)
96
    gpb = None
97
    disable_shared_mem = num_server > 1
98
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
99
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
100
101
102
    num_nodes_to_sample = 202
    batch_size = 32
    train_nid = th.arange(num_nodes_to_sample)
103
104
105
106
107
    dist_graph = DistGraph(
        "test_sampling",
        gpb=gpb,
        part_config=part_config,
    )
108

109
    # Create sampler
Rhett Ying's avatar
Rhett Ying committed
110
    sampler = NeighborSampler(
111
112
113
114
115
        dist_graph,
        [5, 10],
        dgl.distributed.sample_neighbors,
        use_graphbolt=use_graphbolt,
        return_eids=return_eids,
Rhett Ying's avatar
Rhett Ying committed
116
    )
117

118
119
120
    # Enable santity check in distributed sampling.
    os.environ["DGL_DIST_DEBUG"] = "1"

121
122
123
124
    # We need to test creating DistDataLoader multiple times.
    for i in range(2):
        # Create DataLoader for constructing blocks
        dataloader = DistDataLoader(
125
            dataset=train_nid,
126
127
128
            batch_size=batch_size,
            collate_fn=sampler.sample_blocks,
            shuffle=False,
Rhett Ying's avatar
Rhett Ying committed
129
130
            drop_last=drop_last,
        )
131
132
133
134

        groundtruth_g = CitationGraphDataset("cora")[0]
        max_nid = []

135
        for _ in range(2):
Rhett Ying's avatar
Rhett Ying committed
136
137
138
            for idx, blocks in zip(
                range(0, num_nodes_to_sample, batch_size), dataloader
            ):
139
                block = blocks[-1]
Rhett Ying's avatar
Rhett Ying committed
140
                o_src, o_dst = block.edges()
141
142
                src_nodes_id = block.srcdata[dgl.NID][o_src]
                dst_nodes_id = block.dstdata[dgl.NID][o_dst]
143
144
145
146
                max_nid.append(np.max(F.asnumpy(dst_nodes_id)))

                src_nodes_id = orig_nid[src_nodes_id]
                dst_nodes_id = orig_nid[dst_nodes_id]
Rhett Ying's avatar
Rhett Ying committed
147
148
149
                has_edges = groundtruth_g.has_edges_between(
                    src_nodes_id, dst_nodes_id
                )
150
                assert np.all(F.asnumpy(has_edges))
151
152
153
154
155
156
157
158
159
160

                if use_graphbolt and not return_eids:
                    continue
                eids = orig_eid[block.edata[dgl.EID]]
                expected_eids = groundtruth_g.edge_ids(
                    src_nodes_id, dst_nodes_id
                )
                assert th.equal(
                    eids, expected_eids
                ), f"{eids} != {expected_eids}"
161
            if drop_last:
Rhett Ying's avatar
Rhett Ying committed
162
163
164
165
166
167
                assert (
                    np.max(max_nid)
                    == num_nodes_to_sample
                    - 1
                    - num_nodes_to_sample % batch_size
                )
168
169
            else:
                assert np.max(max_nid) == num_nodes_to_sample - 1
170
    del dataloader
Rhett Ying's avatar
Rhett Ying committed
171
172
    # this is needed since there's two test here in one process
    dgl.distributed.exit_client()
173
174


Rhett Ying's avatar
Rhett Ying committed
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
def test_standalone():
    reset_envs()
    with tempfile.TemporaryDirectory() as test_dir:
        ip_config = os.path.join(test_dir, "ip_config.txt")
        generate_ip_config(ip_config, 1, 1)

        g = CitationGraphDataset("cora")[0]
        print(g.idtype)
        num_parts = 1
        num_hops = 1

        orig_nid, orig_eid = partition_graph(
            g,
            "test_sampling",
            num_parts,
            test_dir,
            num_hops=num_hops,
            part_method="metis",
            return_mapping=True,
        )
        part_config = os.path.join(test_dir, "test_sampling.json")
        os.environ["DGL_DIST_MODE"] = "standalone"
        try:
            start_dist_dataloader(
                0, ip_config, part_config, 1, True, orig_nid, orig_eid
            )
        except Exception as e:
            print(e)


def start_dist_neg_dataloader(
    rank,
    ip_config,
    part_config,
    num_server,
    num_workers,
    orig_nid,
    groundtruth_g,
):
214
215
    import dgl
    import torch as th
Rhett Ying's avatar
Rhett Ying committed
216
217

    dgl.distributed.initialize(ip_config)
218
219
220
    gpb = None
    disable_shared_mem = num_server > 1
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
221
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
222
223
    num_edges_to_sample = 202
    batch_size = 32
Rhett Ying's avatar
Rhett Ying committed
224
    dist_graph = DistGraph("test_mp", gpb=gpb, part_config=part_config)
225
226
227
228
229
230
231
232
    assert len(dist_graph.ntypes) == len(groundtruth_g.ntypes)
    assert len(dist_graph.etypes) == len(groundtruth_g.etypes)
    if len(dist_graph.etypes) == 1:
        train_eid = th.arange(num_edges_to_sample)
    else:
        train_eid = {dist_graph.etypes[0]: th.arange(num_edges_to_sample)}

    for i in range(num_server):
Rhett Ying's avatar
Rhett Ying committed
233
        part, _, _, _, _, _, _ = load_partition(part_config, i)
234
235

    num_negs = 5
Rhett Ying's avatar
Rhett Ying committed
236
237
238
239
240
241
242
243
244
245
246
247
    sampler = dgl.dataloading.MultiLayerNeighborSampler([5, 10])
    negative_sampler = dgl.dataloading.negative_sampler.Uniform(num_negs)
    dataloader = dgl.dataloading.DistEdgeDataLoader(
        dist_graph,
        train_eid,
        sampler,
        batch_size=batch_size,
        negative_sampler=negative_sampler,
        shuffle=True,
        drop_last=False,
        num_workers=num_workers,
    )
248
    for _ in range(2):
Rhett Ying's avatar
Rhett Ying committed
249
250
251
        for _, (_, pos_graph, neg_graph, blocks) in zip(
            range(0, num_edges_to_sample, batch_size), dataloader
        ):
252
253
            block = blocks[-1]
            for src_type, etype, dst_type in block.canonical_etypes:
Rhett Ying's avatar
Rhett Ying committed
254
                o_src, o_dst = block.edges(etype=etype)
255
256
257
258
                src_nodes_id = block.srcnodes[src_type].data[dgl.NID][o_src]
                dst_nodes_id = block.dstnodes[dst_type].data[dgl.NID][o_dst]
                src_nodes_id = orig_nid[src_type][src_nodes_id]
                dst_nodes_id = orig_nid[dst_type][dst_nodes_id]
Rhett Ying's avatar
Rhett Ying committed
259
260
261
                has_edges = groundtruth_g.has_edges_between(
                    src_nodes_id, dst_nodes_id, etype=etype
                )
262
                assert np.all(F.asnumpy(has_edges))
Rhett Ying's avatar
Rhett Ying committed
263
264
265
266
267
268
269
270
                assert np.all(
                    F.asnumpy(block.dstnodes[dst_type].data[dgl.NID])
                    == F.asnumpy(pos_graph.nodes[dst_type].data[dgl.NID])
                )
                assert np.all(
                    F.asnumpy(block.dstnodes[dst_type].data[dgl.NID])
                    == F.asnumpy(neg_graph.nodes[dst_type].data[dgl.NID])
                )
271
272
273
                assert pos_graph.num_edges() * num_negs == neg_graph.num_edges()

    del dataloader
Rhett Ying's avatar
Rhett Ying committed
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
    # this is needed since there's two test here in one process
    dgl.distributed.exit_client()


def check_neg_dataloader(g, num_server, num_workers):
    with tempfile.TemporaryDirectory() as test_dir:
        ip_config = "ip_config.txt"
        generate_ip_config(ip_config, num_server, num_server)

        num_parts = num_server
        num_hops = 1
        orig_nid, orig_eid = partition_graph(
            g,
            "test_sampling",
            num_parts,
            test_dir,
            num_hops=num_hops,
            part_method="metis",
            return_mapping=True,
        )
        part_config = os.path.join(test_dir, "test_sampling.json")
        if not isinstance(orig_nid, dict):
            orig_nid = {g.ntypes[0]: orig_nid}
        if not isinstance(orig_eid, dict):
            orig_eid = {g.etypes[0]: orig_eid}

        pserver_list = []
        ctx = mp.get_context("spawn")
        for i in range(num_server):
            p = ctx.Process(
                target=start_server,
                args=(
                    i,
                    ip_config,
                    part_config,
                    num_server > 1,
                    num_workers + 1,
                ),
            )
            p.start()
            time.sleep(1)
            pserver_list.append(p)
        os.environ["DGL_DIST_MODE"] = "distributed"
        os.environ["DGL_NUM_SAMPLER"] = str(num_workers)
        ptrainer_list = []

        p = ctx.Process(
            target=start_dist_neg_dataloader,
            args=(
                0,
                ip_config,
                part_config,
                num_server,
                num_workers,
                orig_nid,
                g,
            ),
        )
332
        p.start()
Rhett Ying's avatar
Rhett Ying committed
333
334
335
336
        ptrainer_list.append(p)

        for p in pserver_list:
            p.join()
337
            assert p.exitcode == 0
Rhett Ying's avatar
Rhett Ying committed
338
339
        for p in ptrainer_list:
            p.join()
340
            assert p.exitcode == 0
Rhett Ying's avatar
Rhett Ying committed
341
342


343
@pytest.mark.parametrize("num_server", [1])
344
@pytest.mark.parametrize("num_workers", [0, 1])
345
346
347
348
349
350
@pytest.mark.parametrize("drop_last", [False, True])
@pytest.mark.parametrize("use_graphbolt", [False, True])
@pytest.mark.parametrize("return_eids", [False, True])
def test_dist_dataloader(
    num_server, num_workers, drop_last, use_graphbolt, return_eids
):
351
    reset_envs()
352
353
    os.environ["DGL_DIST_MODE"] = "distributed"
    os.environ["DGL_NUM_SAMPLER"] = str(num_workers)
Rhett Ying's avatar
Rhett Ying committed
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
    with tempfile.TemporaryDirectory() as test_dir:
        ip_config = "ip_config.txt"
        generate_ip_config(ip_config, num_server, num_server)

        g = CitationGraphDataset("cora")[0]
        num_parts = num_server
        num_hops = 1

        orig_nid, orig_eid = partition_graph(
            g,
            "test_sampling",
            num_parts,
            test_dir,
            num_hops=num_hops,
            part_method="metis",
            return_mapping=True,
370
371
            use_graphbolt=use_graphbolt,
            store_eids=return_eids,
Rhett Ying's avatar
Rhett Ying committed
372
373
374
375
376
377
378
379
380
381
382
383
384
385
        )

        part_config = os.path.join(test_dir, "test_sampling.json")
        pserver_list = []
        ctx = mp.get_context("spawn")
        for i in range(num_server):
            p = ctx.Process(
                target=start_server,
                args=(
                    i,
                    ip_config,
                    part_config,
                    num_server > 1,
                    num_workers + 1,
386
                    use_graphbolt,
Rhett Ying's avatar
Rhett Ying committed
387
388
                ),
            )
389
            p.start()
Rhett Ying's avatar
Rhett Ying committed
390
391
392
393
394
395
            time.sleep(1)
            pserver_list.append(p)

        ptrainer_list = []
        num_trainers = 1
        for trainer_id in range(num_trainers):
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
            p = ctx.Process(
                target=start_dist_dataloader,
                args=(
                    trainer_id,
                    ip_config,
                    part_config,
                    num_server,
                    drop_last,
                    orig_nid,
                    orig_eid,
                    use_graphbolt,
                    return_eids,
                ),
            )
            p.start()
            time.sleep(1)  # avoid race condition when instantiating DistGraph
            ptrainer_list.append(p)
Rhett Ying's avatar
Rhett Ying committed
413
414
415

        for p in ptrainer_list:
            p.join()
416
            assert p.exitcode == 0
417
        for p in pserver_list:
Rhett Ying's avatar
Rhett Ying committed
418
            p.join()
419
            assert p.exitcode == 0
Rhett Ying's avatar
Rhett Ying committed
420
421
422
423
424
425
426
427
428
429
430


def start_node_dataloader(
    rank,
    ip_config,
    part_config,
    num_server,
    num_workers,
    orig_nid,
    orig_eid,
    groundtruth_g,
431
432
    use_graphbolt=False,
    return_eids=False,
Rhett Ying's avatar
Rhett Ying committed
433
434
):
    dgl.distributed.initialize(ip_config)
435
    gpb = None
436
    disable_shared_mem = num_server > 1
437
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
438
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
439
440
    num_nodes_to_sample = 202
    batch_size = 32
441
442
443
444
445
    dist_graph = DistGraph(
        "test_sampling",
        gpb=gpb,
        part_config=part_config,
    )
446
447
448
    assert len(dist_graph.ntypes) == len(groundtruth_g.ntypes)
    assert len(dist_graph.etypes) == len(groundtruth_g.etypes)
    if len(dist_graph.etypes) == 1:
449
        train_nid = th.arange(num_nodes_to_sample, dtype=dist_graph.idtype)
450
    else:
451
452
453
        train_nid = {
            "n3": th.arange(num_nodes_to_sample, dtype=dist_graph.idtype)
        }
454

455
    for i in range(num_server):
Rhett Ying's avatar
Rhett Ying committed
456
        part, _, _, _, _, _, _ = load_partition(part_config, i)
457

458
    # Create sampler
Rhett Ying's avatar
Rhett Ying committed
459
460
461
462
463
464
465
466
467
    sampler = dgl.dataloading.MultiLayerNeighborSampler(
        [
            # test dict for hetero
            {etype: 5 for etype in dist_graph.etypes}
            if len(dist_graph.etypes) > 1
            else 5,
            10,
        ]
    )  # test int for hetero
468

469
470
471
    # Enable santity check in distributed sampling.
    os.environ["DGL_DIST_DEBUG"] = "1"

472
473
474
    # We need to test creating DistDataLoader multiple times.
    for i in range(2):
        # Create DataLoader for constructing blocks
475
        dataloader = dgl.dataloading.DistNodeDataLoader(
476
477
478
479
480
481
            dist_graph,
            train_nid,
            sampler,
            batch_size=batch_size,
            shuffle=True,
            drop_last=False,
Rhett Ying's avatar
Rhett Ying committed
482
483
            num_workers=num_workers,
        )
484

485
        for _ in range(2):
Rhett Ying's avatar
Rhett Ying committed
486
487
488
            for idx, (_, _, blocks) in zip(
                range(0, num_nodes_to_sample, batch_size), dataloader
            ):
489
                block = blocks[-1]
490
491
492
                for c_etype in block.canonical_etypes:
                    src_type, _, dst_type = c_etype
                    o_src, o_dst = block.edges(etype=c_etype)
493
494
495
496
                    src_nodes_id = block.srcnodes[src_type].data[dgl.NID][o_src]
                    dst_nodes_id = block.dstnodes[dst_type].data[dgl.NID][o_dst]
                    src_nodes_id = orig_nid[src_type][src_nodes_id]
                    dst_nodes_id = orig_nid[dst_type][dst_nodes_id]
Rhett Ying's avatar
Rhett Ying committed
497
                    has_edges = groundtruth_g.has_edges_between(
498
                        src_nodes_id, dst_nodes_id, etype=c_etype
Rhett Ying's avatar
Rhett Ying committed
499
                    )
500
                    assert np.all(F.asnumpy(has_edges))
501
502
503

                    if use_graphbolt and not return_eids:
                        continue
504
                    eids = orig_eid[c_etype][block.edges[c_etype].data[dgl.EID]]
505
                    expected_eids = groundtruth_g.edge_ids(
506
                        src_nodes_id, dst_nodes_id, etype=c_etype
507
508
509
510
                    )
                    assert th.equal(
                        eids, expected_eids
                    ), f"{eids} != {expected_eids}"
511
    del dataloader
Rhett Ying's avatar
Rhett Ying committed
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
    # this is needed since there's two test here in one process
    dgl.distributed.exit_client()


def start_edge_dataloader(
    rank,
    ip_config,
    part_config,
    num_server,
    num_workers,
    orig_nid,
    orig_eid,
    groundtruth_g,
):
    dgl.distributed.initialize(ip_config)
527
528
529
    gpb = None
    disable_shared_mem = num_server > 1
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
530
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
531
532
    num_edges_to_sample = 202
    batch_size = 32
533
    dist_graph = DistGraph("test_sampling", gpb=gpb, part_config=part_config)
534
535
536
537
538
539
540
541
    assert len(dist_graph.ntypes) == len(groundtruth_g.ntypes)
    assert len(dist_graph.etypes) == len(groundtruth_g.etypes)
    if len(dist_graph.etypes) == 1:
        train_eid = th.arange(num_edges_to_sample)
    else:
        train_eid = {dist_graph.etypes[0]: th.arange(num_edges_to_sample)}

    for i in range(num_server):
Rhett Ying's avatar
Rhett Ying committed
542
        part, _, _, _, _, _, _ = load_partition(part_config, i)
543
544
545

    # Create sampler
    sampler = dgl.dataloading.MultiLayerNeighborSampler([5, 10])
546

547
548
549
    # We need to test creating DistDataLoader multiple times.
    for i in range(2):
        # Create DataLoader for constructing blocks
550
        dataloader = dgl.dataloading.DistEdgeDataLoader(
551
552
553
554
555
556
            dist_graph,
            train_eid,
            sampler,
            batch_size=batch_size,
            shuffle=True,
            drop_last=False,
Rhett Ying's avatar
Rhett Ying committed
557
558
            num_workers=num_workers,
        )
559
560

        for epoch in range(2):
Rhett Ying's avatar
Rhett Ying committed
561
562
563
            for idx, (input_nodes, pos_pair_graph, blocks) in zip(
                range(0, num_edges_to_sample, batch_size), dataloader
            ):
564
565
                block = blocks[-1]
                for src_type, etype, dst_type in block.canonical_etypes:
Rhett Ying's avatar
Rhett Ying committed
566
                    o_src, o_dst = block.edges(etype=etype)
567
568
569
570
                    src_nodes_id = block.srcnodes[src_type].data[dgl.NID][o_src]
                    dst_nodes_id = block.dstnodes[dst_type].data[dgl.NID][o_dst]
                    src_nodes_id = orig_nid[src_type][src_nodes_id]
                    dst_nodes_id = orig_nid[dst_type][dst_nodes_id]
Rhett Ying's avatar
Rhett Ying committed
571
572
573
                    has_edges = groundtruth_g.has_edges_between(
                        src_nodes_id, dst_nodes_id, etype=etype
                    )
574
                    assert np.all(F.asnumpy(has_edges))
Rhett Ying's avatar
Rhett Ying committed
575
576
577
578
579
580
                    assert np.all(
                        F.asnumpy(block.dstnodes[dst_type].data[dgl.NID])
                        == F.asnumpy(
                            pos_pair_graph.nodes[dst_type].data[dgl.NID]
                        )
                    )
581
    del dataloader
Rhett Ying's avatar
Rhett Ying committed
582
583
584
    dgl.distributed.exit_client()


585
586
587
588
589
590
591
592
def check_dataloader(
    g,
    num_server,
    num_workers,
    dataloader_type,
    use_graphbolt=False,
    return_eids=False,
):
Rhett Ying's avatar
Rhett Ying committed
593
594
595
596
597
598
599
600
601
602
603
604
605
606
    with tempfile.TemporaryDirectory() as test_dir:
        ip_config = "ip_config.txt"
        generate_ip_config(ip_config, num_server, num_server)

        num_parts = num_server
        num_hops = 1
        orig_nid, orig_eid = partition_graph(
            g,
            "test_sampling",
            num_parts,
            test_dir,
            num_hops=num_hops,
            part_method="metis",
            return_mapping=True,
607
608
            use_graphbolt=use_graphbolt,
            store_eids=return_eids,
Rhett Ying's avatar
Rhett Ying committed
609
610
611
612
613
        )
        part_config = os.path.join(test_dir, "test_sampling.json")
        if not isinstance(orig_nid, dict):
            orig_nid = {g.ntypes[0]: orig_nid}
        if not isinstance(orig_eid, dict):
614
            orig_eid = {g.canonical_etypes[0]: orig_eid}
Rhett Ying's avatar
Rhett Ying committed
615
616
617
618
619
620
621
622
623
624
625
626

        pserver_list = []
        ctx = mp.get_context("spawn")
        for i in range(num_server):
            p = ctx.Process(
                target=start_server,
                args=(
                    i,
                    ip_config,
                    part_config,
                    num_server > 1,
                    num_workers + 1,
627
                    use_graphbolt,
Rhett Ying's avatar
Rhett Ying committed
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
                ),
            )
            p.start()
            time.sleep(1)
            pserver_list.append(p)

        os.environ["DGL_DIST_MODE"] = "distributed"
        os.environ["DGL_NUM_SAMPLER"] = str(num_workers)
        ptrainer_list = []
        if dataloader_type == "node":
            p = ctx.Process(
                target=start_node_dataloader,
                args=(
                    0,
                    ip_config,
                    part_config,
                    num_server,
                    num_workers,
                    orig_nid,
                    orig_eid,
                    g,
649
650
                    use_graphbolt,
                    return_eids,
Rhett Ying's avatar
Rhett Ying committed
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
                ),
            )
            p.start()
            ptrainer_list.append(p)
        elif dataloader_type == "edge":
            p = ctx.Process(
                target=start_edge_dataloader,
                args=(
                    0,
                    ip_config,
                    part_config,
                    num_server,
                    num_workers,
                    orig_nid,
                    orig_eid,
                    g,
                ),
            )
            p.start()
            ptrainer_list.append(p)
        for p in pserver_list:
            p.join()
673
            assert p.exitcode == 0
Rhett Ying's avatar
Rhett Ying committed
674
675
        for p in ptrainer_list:
            p.join()
676
            assert p.exitcode == 0
Rhett Ying's avatar
Rhett Ying committed
677

678

679
def create_random_hetero():
Rhett Ying's avatar
Rhett Ying committed
680
681
    num_nodes = {"n1": 10000, "n2": 10010, "n3": 10020}
    etypes = [("n1", "r1", "n2"), ("n1", "r2", "n3"), ("n2", "r3", "n3")]
682
683
684
    edges = {}
    for etype in etypes:
        src_ntype, _, dst_ntype = etype
Rhett Ying's avatar
Rhett Ying committed
685
686
687
688
689
690
691
        arr = spsp.random(
            num_nodes[src_ntype],
            num_nodes[dst_ntype],
            density=0.001,
            format="coo",
            random_state=100,
        )
692
693
        edges[etype] = (arr.row, arr.col)
    g = dgl.heterograph(edges, num_nodes)
Hongzhi (Steve), Chen's avatar
Hongzhi (Steve), Chen committed
694
695
    g.nodes["n1"].data["feat"] = F.unsqueeze(F.arange(0, g.num_nodes("n1")), 1)
    g.edges["r1"].data["feat"] = F.unsqueeze(F.arange(0, g.num_edges("r1")), 1)
696
697
    return g

Rhett Ying's avatar
Rhett Ying committed
698

699
700
@pytest.mark.parametrize("num_server", [1])
@pytest.mark.parametrize("num_workers", [0, 1])
701
@pytest.mark.parametrize("dataloader_type", ["node", "edge"])
702
703
704
705
706
707
708
709
@pytest.mark.parametrize("use_graphbolt", [False, True])
@pytest.mark.parametrize("return_eids", [False, True])
def test_dataloader_homograph(
    num_server, num_workers, dataloader_type, use_graphbolt, return_eids
):
    if dataloader_type == "edge" and use_graphbolt:
        # GraphBolt does not support edge dataloader.
        return
710
    reset_envs()
711
    g = CitationGraphDataset("cora")[0]
712
713
714
715
716
717
718
719
720
721
722
723
724
    check_dataloader(
        g,
        num_server,
        num_workers,
        dataloader_type,
        use_graphbolt=use_graphbolt,
        return_eids=return_eids,
    )


@pytest.mark.parametrize("num_server", [1])
@pytest.mark.parametrize("num_workers", [0, 1])
@pytest.mark.parametrize("dataloader_type", ["node", "edge"])
725
726
727
728
729
730
731
732
@pytest.mark.parametrize("use_graphbolt", [False, True])
@pytest.mark.parametrize("return_eids", [False, True])
def test_dataloader_heterograph(
    num_server, num_workers, dataloader_type, use_graphbolt, return_eids
):
    if dataloader_type == "edge" and use_graphbolt:
        # GraphBolt does not support edge dataloader.
        return
733
    reset_envs()
734
    g = create_random_hetero()
735
736
737
738
739
740
741
742
    check_dataloader(
        g,
        num_server,
        num_workers,
        dataloader_type,
        use_graphbolt=use_graphbolt,
        return_eids=return_eids,
    )
Rhett Ying's avatar
Rhett Ying committed
743

744

745
@unittest.skip(reason="Skip due to glitch in CI")
746
747
@pytest.mark.parametrize("num_server", [3])
@pytest.mark.parametrize("num_workers", [0, 4])
Rhett Ying's avatar
Rhett Ying committed
748
def test_neg_dataloader(num_server, num_workers):
749
    reset_envs()
750
    g = CitationGraphDataset("cora")[0]
Rhett Ying's avatar
Rhett Ying committed
751
    check_neg_dataloader(g, num_server, num_workers)
752
    g = create_random_hetero()
Rhett Ying's avatar
Rhett Ying committed
753
754
755
756
    check_neg_dataloader(g, num_server, num_workers)


def start_multiple_dataloaders(
757
758
759
760
761
762
763
    ip_config,
    part_config,
    graph_name,
    orig_g,
    num_dataloaders,
    dataloader_type,
    use_graphbolt,
Rhett Ying's avatar
Rhett Ying committed
764
765
):
    dgl.distributed.initialize(ip_config)
766
    dist_g = dgl.distributed.DistGraph(graph_name, part_config=part_config)
Rhett Ying's avatar
Rhett Ying committed
767
    if dataloader_type == "node":
768
        train_ids = th.arange(orig_g.num_nodes(), dtype=dist_g.idtype)
Rhett Ying's avatar
Rhett Ying committed
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
        batch_size = orig_g.num_nodes() // 100
    else:
        train_ids = th.arange(orig_g.num_edges())
        batch_size = orig_g.num_edges() // 100
    sampler = dgl.dataloading.NeighborSampler([-1])
    dataloaders = []
    dl_iters = []
    for _ in range(num_dataloaders):
        if dataloader_type == "node":
            dataloader = dgl.dataloading.DistNodeDataLoader(
                dist_g, train_ids, sampler, batch_size=batch_size
            )
        else:
            dataloader = dgl.dataloading.DistEdgeDataLoader(
                dist_g, train_ids, sampler, batch_size=batch_size
            )
        dataloaders.append(dataloader)
        dl_iters.append(iter(dataloader))

    # iterate on multiple dataloaders randomly
    while len(dl_iters) > 0:
        next_dl = np.random.choice(len(dl_iters), 1)[0]
        try:
            _ = next(dl_iters[next_dl])
        except StopIteration:
            dl_iters.pop(next_dl)
            del dataloaders[next_dl]

    dgl.distributed.exit_client()


@pytest.mark.parametrize("num_dataloaders", [1, 4])
801
@pytest.mark.parametrize("num_workers", [0, 1])
Rhett Ying's avatar
Rhett Ying committed
802
@pytest.mark.parametrize("dataloader_type", ["node", "edge"])
803
804
@pytest.mark.parametrize("use_graphbolt", [False, True])
@pytest.mark.parametrize("return_eids", [False, True])
Rhett Ying's avatar
Rhett Ying committed
805
def test_multiple_dist_dataloaders(
806
    num_dataloaders, num_workers, dataloader_type, use_graphbolt, return_eids
Rhett Ying's avatar
Rhett Ying committed
807
):
808
809
810
    if dataloader_type == "edge" and use_graphbolt:
        # GraphBolt does not support edge dataloader.
        return
Rhett Ying's avatar
Rhett Ying committed
811
812
813
814
815
816
817
818
819
820
    reset_envs()
    os.environ["DGL_DIST_MODE"] = "distributed"
    os.environ["DGL_NUM_SAMPLER"] = str(num_workers)
    num_parts = 1
    num_servers = 1
    with tempfile.TemporaryDirectory() as test_dir:
        ip_config = os.path.join(test_dir, "ip_config.txt")
        generate_ip_config(ip_config, num_parts, num_servers)

        orig_g = dgl.rand_graph(1000, 10000)
821
822
823
824
825
826
827
828
829
        graph_name = "test_multiple_dataloaders"
        partition_graph(
            orig_g,
            graph_name,
            num_parts,
            test_dir,
            use_graphbolt=use_graphbolt,
            store_eids=return_eids,
        )
Rhett Ying's avatar
Rhett Ying committed
830
831
832
833
834
835
836
837
838
839
840
841
842
        part_config = os.path.join(test_dir, f"{graph_name}.json")

        p_servers = []
        ctx = mp.get_context("spawn")
        for i in range(num_servers):
            p = ctx.Process(
                target=start_server,
                args=(
                    i,
                    ip_config,
                    part_config,
                    num_servers > 1,
                    num_workers + 1,
843
                    use_graphbolt,
Rhett Ying's avatar
Rhett Ying committed
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
                ),
            )
            p.start()
            time.sleep(1)
            p_servers.append(p)

        p_client = ctx.Process(
            target=start_multiple_dataloaders,
            args=(
                ip_config,
                part_config,
                graph_name,
                orig_g,
                num_dataloaders,
                dataloader_type,
859
                use_graphbolt,
Rhett Ying's avatar
Rhett Ying committed
860
861
862
863
864
            ),
        )
        p_client.start()

        p_client.join()
865
        assert p_client.exitcode == 0
Rhett Ying's avatar
Rhett Ying committed
866
867
        for p in p_servers:
            p.join()
868
            assert p.exitcode == 0
Rhett Ying's avatar
Rhett Ying committed
869
    reset_envs()