test_mp_dataloader.py 23.7 KB
Newer Older
1
import multiprocessing as mp
Rhett Ying's avatar
Rhett Ying committed
2
3
import os
import tempfile
4
import time
Rhett Ying's avatar
Rhett Ying committed
5

6
import backend as F
Rhett Ying's avatar
Rhett Ying committed
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
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

22
23
24
25
26
27
28
29
30

class NeighborSampler(object):
    def __init__(self, g, fanouts, sample_neighbors):
        self.g = g
        self.fanouts = fanouts
        self.sample_neighbors = sample_neighbors

    def sample_blocks(self, seeds):
        import torch as th
Rhett Ying's avatar
Rhett Ying committed
31

32
33
34
35
36
        seeds = th.LongTensor(np.asarray(seeds))
        blocks = []
        for fanout in self.fanouts:
            # For each seed node, sample ``fanout`` neighbors.
            frontier = self.sample_neighbors(
Rhett Ying's avatar
Rhett Ying committed
37
38
39
40
                self.g, seeds, fanout, replace=True
            )
            # Then we compact the frontier into a bipartite graph for
            # message passing.
41
42
43
44
45
46
47
48
            block = dgl.to_block(frontier, seeds)
            # Obtain the seed nodes for next layer.
            seeds = block.srcdata[dgl.NID]

            blocks.insert(0, block)
        return blocks


Rhett Ying's avatar
Rhett Ying committed
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
def start_server(
    rank,
    ip_config,
    part_config,
    disable_shared_mem,
    num_clients,
    keep_alive=False,
):
    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"],
        keep_alive=keep_alive,
    )
68
69
70
    g.start()


Rhett Ying's avatar
Rhett Ying committed
71
72
73
74
75
76
77
78
79
80
def start_dist_dataloader(
    rank,
    ip_config,
    part_config,
    num_server,
    drop_last,
    orig_nid,
    orig_eid,
    group_id=0,
):
81
82
    import dgl
    import torch as th
Rhett Ying's avatar
Rhett Ying committed
83
84
85

    os.environ["DGL_GROUP_ID"] = str(group_id)
    dgl.distributed.initialize(ip_config)
86
    gpb = None
87
    disable_shared_mem = num_server > 0
88
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
89
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
90
91
92
    num_nodes_to_sample = 202
    batch_size = 32
    train_nid = th.arange(num_nodes_to_sample)
Rhett Ying's avatar
Rhett Ying committed
93
    dist_graph = DistGraph("test_mp", gpb=gpb, part_config=part_config)
94

95
    for i in range(num_server):
Rhett Ying's avatar
Rhett Ying committed
96
        part, _, _, _, _, _, _ = load_partition(part_config, i)
97

98
    # Create sampler
Rhett Ying's avatar
Rhett Ying committed
99
100
101
    sampler = NeighborSampler(
        dist_graph, [5, 10], dgl.distributed.sample_neighbors
    )
102

103
104
105
106
107
108
109
110
    # We need to test creating DistDataLoader multiple times.
    for i in range(2):
        # Create DataLoader for constructing blocks
        dataloader = DistDataLoader(
            dataset=train_nid.numpy(),
            batch_size=batch_size,
            collate_fn=sampler.sample_blocks,
            shuffle=False,
Rhett Ying's avatar
Rhett Ying committed
111
112
            drop_last=drop_last,
        )
113
114
115
116
117

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

        for epoch in range(2):
Rhett Ying's avatar
Rhett Ying committed
118
119
120
            for idx, blocks in zip(
                range(0, num_nodes_to_sample, batch_size), dataloader
            ):
121
                block = blocks[-1]
Rhett Ying's avatar
Rhett Ying committed
122
                o_src, o_dst = block.edges()
123
124
                src_nodes_id = block.srcdata[dgl.NID][o_src]
                dst_nodes_id = block.dstdata[dgl.NID][o_dst]
125
126
127
128
                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
129
130
131
                has_edges = groundtruth_g.has_edges_between(
                    src_nodes_id, dst_nodes_id
                )
132
133
                assert np.all(F.asnumpy(has_edges))
            if drop_last:
Rhett Ying's avatar
Rhett Ying committed
134
135
136
137
138
139
                assert (
                    np.max(max_nid)
                    == num_nodes_to_sample
                    - 1
                    - num_nodes_to_sample % batch_size
                )
140
141
            else:
                assert np.max(max_nid) == num_nodes_to_sample - 1
142
    del dataloader
Rhett Ying's avatar
Rhett Ying committed
143
144
    # this is needed since there's two test here in one process
    dgl.distributed.exit_client()
145
146


Rhett Ying's avatar
Rhett Ying committed
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
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",
            reshuffle=True,
            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,
):
187
188
    import dgl
    import torch as th
Rhett Ying's avatar
Rhett Ying committed
189
190

    dgl.distributed.initialize(ip_config)
191
192
193
    gpb = None
    disable_shared_mem = num_server > 1
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
194
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
195
196
    num_edges_to_sample = 202
    batch_size = 32
Rhett Ying's avatar
Rhett Ying committed
197
    dist_graph = DistGraph("test_mp", gpb=gpb, part_config=part_config)
198
199
200
201
202
203
204
205
    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
206
        part, _, _, _, _, _, _ = load_partition(part_config, i)
207
208

    num_negs = 5
Rhett Ying's avatar
Rhett Ying committed
209
210
211
212
213
214
215
216
217
218
219
220
    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,
    )
221
    for _ in range(2):
Rhett Ying's avatar
Rhett Ying committed
222
223
224
        for _, (_, pos_graph, neg_graph, blocks) in zip(
            range(0, num_edges_to_sample, batch_size), dataloader
        ):
225
226
            block = blocks[-1]
            for src_type, etype, dst_type in block.canonical_etypes:
Rhett Ying's avatar
Rhett Ying committed
227
                o_src, o_dst = block.edges(etype=etype)
228
229
230
231
                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
232
233
234
                has_edges = groundtruth_g.has_edges_between(
                    src_nodes_id, dst_nodes_id, etype=etype
                )
235
                assert np.all(F.asnumpy(has_edges))
Rhett Ying's avatar
Rhett Ying committed
236
237
238
239
240
241
242
243
                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])
                )
244
245
246
                assert pos_graph.num_edges() * num_negs == neg_graph.num_edges()

    del dataloader
Rhett Ying's avatar
Rhett Ying committed
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
    # 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",
            reshuffle=True,
            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,
            ),
        )
306
        p.start()
Rhett Ying's avatar
Rhett Ying committed
307
308
309
310
311
312
313
314
        ptrainer_list.append(p)

        for p in pserver_list:
            p.join()
        for p in ptrainer_list:
            p.join()


315
@pytest.mark.parametrize("num_server", [3])
316
@pytest.mark.parametrize("num_workers", [0, 4])
317
@pytest.mark.parametrize("drop_last", [True, False])
318
@pytest.mark.parametrize("reshuffle", [True, False])
319
@pytest.mark.parametrize("num_groups", [1])
Rhett Ying's avatar
Rhett Ying committed
320
321
322
def test_dist_dataloader(
    num_server, num_workers, drop_last, reshuffle, num_groups
):
323
    reset_envs()
324
325
326
327
    # No multiple partitions on single machine for
    # multiple client groups in case of race condition.
    if num_groups > 1:
        num_server = 1
Rhett Ying's avatar
Rhett Ying committed
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
    with tempfile.TemporaryDirectory() as test_dir:
        ip_config = "ip_config.txt"
        generate_ip_config(ip_config, num_server, num_server)

        g = CitationGraphDataset("cora")[0]
        print(g.idtype)
        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",
            reshuffle=reshuffle,
            return_mapping=True,
        )

        part_config = os.path.join(test_dir, "test_sampling.json")
        pserver_list = []
        ctx = mp.get_context("spawn")
        keep_alive = num_groups > 1
        for i in range(num_server):
            p = ctx.Process(
                target=start_server,
                args=(
                    i,
                    ip_config,
                    part_config,
                    num_server > 1,
                    num_workers + 1,
                    keep_alive,
                ),
            )
364
            p.start()
Rhett Ying's avatar
Rhett Ying committed
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
            time.sleep(1)
            pserver_list.append(p)

        os.environ["DGL_DIST_MODE"] = "distributed"
        os.environ["DGL_NUM_SAMPLER"] = str(num_workers)
        ptrainer_list = []
        num_trainers = 1
        for trainer_id in range(num_trainers):
            for group_id in range(num_groups):
                p = ctx.Process(
                    target=start_dist_dataloader,
                    args=(
                        trainer_id,
                        ip_config,
                        part_config,
                        num_server,
                        drop_last,
                        orig_nid,
                        orig_eid,
                        group_id,
                    ),
                )
                p.start()
                time.sleep(
                    1
                )  # avoid race condition when instantiating DistGraph
                ptrainer_list.append(p)

        for p in ptrainer_list:
            p.join()
        if keep_alive:
            for p in pserver_list:
                assert p.is_alive()
            # force shutdown server
            dgl.distributed.shutdown_servers("mp_ip_config.txt", 1)
400
        for p in pserver_list:
Rhett Ying's avatar
Rhett Ying committed
401
402
403
404
405
406
407
408
409
410
411
412
413
414
            p.join()


def start_node_dataloader(
    rank,
    ip_config,
    part_config,
    num_server,
    num_workers,
    orig_nid,
    orig_eid,
    groundtruth_g,
):
    dgl.distributed.initialize(ip_config)
415
    gpb = None
416
    disable_shared_mem = num_server > 1
417
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
418
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
419
420
    num_nodes_to_sample = 202
    batch_size = 32
Rhett Ying's avatar
Rhett Ying committed
421
    dist_graph = DistGraph("test_mp", gpb=gpb, part_config=part_config)
422
423
424
425
426
    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_nid = th.arange(num_nodes_to_sample)
    else:
Rhett Ying's avatar
Rhett Ying committed
427
        train_nid = {"n3": th.arange(num_nodes_to_sample)}
428

429
    for i in range(num_server):
Rhett Ying's avatar
Rhett Ying committed
430
        part, _, _, _, _, _, _ = load_partition(part_config, i)
431

432
    # Create sampler
Rhett Ying's avatar
Rhett Ying committed
433
434
435
436
437
438
439
440
441
    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
442
443
444
445

    # We need to test creating DistDataLoader multiple times.
    for i in range(2):
        # Create DataLoader for constructing blocks
446
        dataloader = dgl.dataloading.DistNodeDataLoader(
447
448
449
450
451
452
            dist_graph,
            train_nid,
            sampler,
            batch_size=batch_size,
            shuffle=True,
            drop_last=False,
Rhett Ying's avatar
Rhett Ying committed
453
454
            num_workers=num_workers,
        )
455
456

        for epoch in range(2):
Rhett Ying's avatar
Rhett Ying committed
457
458
459
            for idx, (_, _, blocks) in zip(
                range(0, num_nodes_to_sample, batch_size), dataloader
            ):
460
                block = blocks[-1]
461
                for src_type, etype, dst_type in block.canonical_etypes:
Rhett Ying's avatar
Rhett Ying committed
462
                    o_src, o_dst = block.edges(etype=etype)
463
464
465
466
                    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
467
468
469
                    has_edges = groundtruth_g.has_edges_between(
                        src_nodes_id, dst_nodes_id, etype=etype
                    )
470
                    assert np.all(F.asnumpy(has_edges))
471
    del dataloader
Rhett Ying's avatar
Rhett Ying committed
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
    # 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)
487
488
489
    gpb = None
    disable_shared_mem = num_server > 1
    if disable_shared_mem:
Rhett Ying's avatar
Rhett Ying committed
490
        _, _, _, gpb, _, _, _ = load_partition(part_config, rank)
491
492
    num_edges_to_sample = 202
    batch_size = 32
Rhett Ying's avatar
Rhett Ying committed
493
    dist_graph = DistGraph("test_mp", gpb=gpb, part_config=part_config)
494
495
496
497
498
499
500
501
    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
502
        part, _, _, _, _, _, _ = load_partition(part_config, i)
503
504
505

    # Create sampler
    sampler = dgl.dataloading.MultiLayerNeighborSampler([5, 10])
506

507
508
509
    # We need to test creating DistDataLoader multiple times.
    for i in range(2):
        # Create DataLoader for constructing blocks
510
        dataloader = dgl.dataloading.DistEdgeDataLoader(
511
512
513
514
515
516
            dist_graph,
            train_eid,
            sampler,
            batch_size=batch_size,
            shuffle=True,
            drop_last=False,
Rhett Ying's avatar
Rhett Ying committed
517
518
            num_workers=num_workers,
        )
519
520

        for epoch in range(2):
Rhett Ying's avatar
Rhett Ying committed
521
522
523
            for idx, (input_nodes, pos_pair_graph, blocks) in zip(
                range(0, num_edges_to_sample, batch_size), dataloader
            ):
524
525
                block = blocks[-1]
                for src_type, etype, dst_type in block.canonical_etypes:
Rhett Ying's avatar
Rhett Ying committed
526
                    o_src, o_dst = block.edges(etype=etype)
527
528
529
530
                    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
531
532
533
                    has_edges = groundtruth_g.has_edges_between(
                        src_nodes_id, dst_nodes_id, etype=etype
                    )
534
                    assert np.all(F.asnumpy(has_edges))
Rhett Ying's avatar
Rhett Ying committed
535
536
537
538
539
540
                    assert np.all(
                        F.asnumpy(block.dstnodes[dst_type].data[dgl.NID])
                        == F.asnumpy(
                            pos_pair_graph.nodes[dst_type].data[dgl.NID]
                        )
                    )
541
    del dataloader
Rhett Ying's avatar
Rhett Ying committed
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
    dgl.distributed.exit_client()


def check_dataloader(g, num_server, num_workers, dataloader_type):
    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",
            reshuffle=True,
            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 = []
        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,
                ),
            )
            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()
        for p in ptrainer_list:
            p.join()

625

626
def create_random_hetero():
Rhett Ying's avatar
Rhett Ying committed
627
628
    num_nodes = {"n1": 10000, "n2": 10010, "n3": 10020}
    etypes = [("n1", "r1", "n2"), ("n1", "r2", "n3"), ("n2", "r3", "n3")]
629
630
631
    edges = {}
    for etype in etypes:
        src_ntype, _, dst_ntype = etype
Rhett Ying's avatar
Rhett Ying committed
632
633
634
635
636
637
638
        arr = spsp.random(
            num_nodes[src_ntype],
            num_nodes[dst_ntype],
            density=0.001,
            format="coo",
            random_state=100,
        )
639
640
        edges[etype] = (arr.row, arr.col)
    g = dgl.heterograph(edges, num_nodes)
Rhett Ying's avatar
Rhett Ying committed
641
642
643
644
645
646
    g.nodes["n1"].data["feat"] = F.unsqueeze(
        F.arange(0, g.number_of_nodes("n1")), 1
    )
    g.edges["r1"].data["feat"] = F.unsqueeze(
        F.arange(0, g.number_of_edges("r1")), 1
    )
647
648
    return g

Rhett Ying's avatar
Rhett Ying committed
649

650
651
652
@pytest.mark.parametrize("num_server", [3])
@pytest.mark.parametrize("num_workers", [0, 4])
@pytest.mark.parametrize("dataloader_type", ["node", "edge"])
Rhett Ying's avatar
Rhett Ying committed
653
def test_dataloader(num_server, num_workers, dataloader_type):
654
    reset_envs()
655
    g = CitationGraphDataset("cora")[0]
Rhett Ying's avatar
Rhett Ying committed
656
    check_dataloader(g, num_server, num_workers, dataloader_type)
657
    g = create_random_hetero()
Rhett Ying's avatar
Rhett Ying committed
658
659
    check_dataloader(g, num_server, num_workers, dataloader_type)

660

661
662
@pytest.mark.parametrize("num_server", [3])
@pytest.mark.parametrize("num_workers", [0, 4])
Rhett Ying's avatar
Rhett Ying committed
663
def test_neg_dataloader(num_server, num_workers):
664
    reset_envs()
665
    g = CitationGraphDataset("cora")[0]
Rhett Ying's avatar
Rhett Ying committed
666
    check_neg_dataloader(g, num_server, num_workers)
667
    g = create_random_hetero()
Rhett Ying's avatar
Rhett Ying committed
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
    check_neg_dataloader(g, num_server, num_workers)


def start_multiple_dataloaders(
    ip_config, part_config, graph_name, orig_g, num_dataloaders, dataloader_type
):
    dgl.distributed.initialize(ip_config)
    dist_g = dgl.distributed.DistGraph(graph_name, part_config=part_config)
    if dataloader_type == "node":
        train_ids = th.arange(orig_g.num_nodes())
        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])
@pytest.mark.parametrize("num_workers", [0, 1, 4])
@pytest.mark.parametrize("dataloader_type", ["node", "edge"])
def test_multiple_dist_dataloaders(
    num_dataloaders, num_workers, dataloader_type
):
    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)
        graph_name = "test"
        partition_graph(orig_g, graph_name, num_parts, test_dir)
        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,
                ),
            )
            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,
            ),
        )
        p_client.start()

        p_client.join()
        for p in p_servers:
            p.join()
    reset_envs()