File size: 23,543 Bytes
0cb481f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Compute-node smoke test for :class:`CompactGraphDataset`.

The test intentionally samples rather than scans the complete dataset.  It
exercises:

* deterministic random graph reconstruction;
* at least one graph from every selected shard and cross-shard batches;
* the per-process mmap shard cache through repeated access;
* PyG DataLoader collation with zero and multiple worker processes; and
* model-facing tensor shapes, dtypes, index ranges, and finite values.

Timing, process RSS/high-water marks, page faults, and filesystem I/O counters
are written to an atomic JSON report.  GNU ``time -v`` and Slurm accounting in
the companion sbatch file provide job-wide measurements including workers.
"""

from __future__ import annotations

import argparse
import gc
import json
import math
import os
import random
import resource
import socket
import statistics
import sys
import time
import traceback
from pathlib import Path
from typing import Any, Dict, List, Mapping, Sequence

import torch
import torch_geometric
from torch.utils.data import Subset
from torch_geometric.data import Batch, Data
from torch_geometric.loader import DataLoader

from compact_graph_dataset import CompactGraphDataset


MIB = 1024 * 1024


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Sample, batch, and profile a compact GNNCP dataset."
    )
    parser.add_argument("--compact", required=True, help="Compact dataset directory")
    parser.add_argument("--report", help="Atomic JSON report path")
    parser.add_argument("--num-random", type=int, default=16)
    parser.add_argument(
        "--max-shards",
        type=int,
        default=0,
        help="Maximum shards to probe; 0 tests every shard",
    )
    parser.add_argument("--batch-size", type=int, default=4)
    parser.add_argument("--num-workers", type=int, default=2)
    parser.add_argument(
        "--worker-timeout-s",
        type=float,
        default=180.0,
        help="Multi-worker DataLoader timeout; zero disables it",
    )
    parser.add_argument("--max-batches", type=int, default=8)
    parser.add_argument("--max-cached-shards", type=int, default=2)
    parser.add_argument("--repeat-count", type=int, default=3)
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument(
        "--torch-threads",
        type=int,
        default=min(4, int(os.environ.get("SLURM_CPUS_PER_TASK", "4"))),
    )
    parser.add_argument(
        "--no-strict",
        action="store_true",
        help="Disable loader invariant checks (not recommended for smoke tests)",
    )
    args = parser.parse_args()
    positive = {
        "num_random": args.num_random,
        "batch_size": args.batch_size,
        "max_batches": args.max_batches,
        "max_cached_shards": args.max_cached_shards,
        "repeat_count": args.repeat_count,
        "torch_threads": args.torch_threads,
    }
    for name, value in positive.items():
        if value < 1:
            parser.error(f"--{name.replace('_', '-')} must be >= 1")
    if args.num_workers < 0:
        parser.error("--num-workers must be >= 0")
    if args.worker_timeout_s < 0:
        parser.error("--worker-timeout-s must be >= 0")
    if args.max_shards < 0:
        parser.error("--max-shards must be >= 0")
    return args


def _proc_status_mib(field: str) -> float | None:
    try:
        with Path("/proc/self/status").open("r", encoding="utf-8") as handle:
            for line in handle:
                if line.startswith(f"{field}:"):
                    return float(line.split()[1]) / 1024.0
    except OSError:
        return None
    return None


def _proc_io_bytes() -> Dict[str, int]:
    result = {"read_bytes": 0, "write_bytes": 0}
    try:
        with Path("/proc/self/io").open("r", encoding="utf-8") as handle:
            for line in handle:
                key, raw_value = line.split(":", 1)
                if key in result:
                    result[key] = int(raw_value.strip())
    except OSError:
        pass
    return result


def resource_snapshot() -> Dict[str, float | int | None]:
    usage = resource.getrusage(resource.RUSAGE_SELF)
    children = resource.getrusage(resource.RUSAGE_CHILDREN)
    io_bytes = _proc_io_bytes()
    # ru_maxrss is KiB on Linux, which is the target Slurm platform.
    return {
        "monotonic_s": time.perf_counter(),
        "rss_mib": _proc_status_mib("VmRSS"),
        "hwm_mib": _proc_status_mib("VmHWM"),
        "ru_maxrss_mib": float(usage.ru_maxrss) / 1024.0,
        "minor_faults": int(usage.ru_minflt),
        "major_faults": int(usage.ru_majflt),
        "self_user_cpu_s": float(usage.ru_utime),
        "self_system_cpu_s": float(usage.ru_stime),
        "children_ru_maxrss_mib": float(children.ru_maxrss) / 1024.0,
        "children_minor_faults": int(children.ru_minflt),
        "children_major_faults": int(children.ru_majflt),
        "children_user_cpu_s": float(children.ru_utime),
        "children_system_cpu_s": float(children.ru_stime),
        "read_bytes": io_bytes["read_bytes"],
        "write_bytes": io_bytes["write_bytes"],
    }


def resource_delta(
    before: Mapping[str, float | int | None],
    after: Mapping[str, float | int | None],
) -> Dict[str, float | int | None]:
    def subtract(key: str) -> float | int | None:
        left = after.get(key)
        right = before.get(key)
        if left is None or right is None:
            return None
        return left - right

    return {
        "elapsed_s": subtract("monotonic_s"),
        "rss_mib_after": after.get("rss_mib"),
        "rss_mib_delta": subtract("rss_mib"),
        "hwm_mib_after": after.get("hwm_mib"),
        "ru_maxrss_mib_after": after.get("ru_maxrss_mib"),
        "minor_faults_delta": subtract("minor_faults"),
        "major_faults_delta": subtract("major_faults"),
        "self_user_cpu_s_delta": subtract("self_user_cpu_s"),
        "self_system_cpu_s_delta": subtract("self_system_cpu_s"),
        "children_ru_maxrss_mib_after": after.get("children_ru_maxrss_mib"),
        "children_minor_faults_delta": subtract("children_minor_faults"),
        "children_major_faults_delta": subtract("children_major_faults"),
        "children_user_cpu_s_delta": subtract("children_user_cpu_s"),
        "children_system_cpu_s_delta": subtract("children_system_cpu_s"),
        "read_mib_delta": (
            None
            if subtract("read_bytes") is None
            else float(subtract("read_bytes")) / MIB
        ),
        "write_mib_delta": (
            None
            if subtract("write_bytes") is None
            else float(subtract("write_bytes")) / MIB
        ),
    }


def latency_summary(values: Sequence[float]) -> Dict[str, float | int]:
    if not values:
        return {"count": 0}
    ordered = sorted(values)
    p95_index = max(0, math.ceil(0.95 * len(ordered)) - 1)
    return {
        "count": len(ordered),
        "total_s": float(sum(ordered)),
        "mean_ms": float(statistics.fmean(ordered) * 1000.0),
        "median_ms": float(statistics.median(ordered) * 1000.0),
        "p95_ms": float(ordered[p95_index] * 1000.0),
        "min_ms": float(ordered[0] * 1000.0),
        "max_ms": float(ordered[-1] * 1000.0),
    }


def ensure_finite(name: str, tensor: torch.Tensor) -> None:
    if not bool(torch.isfinite(tensor).all().item()):
        raise RuntimeError(f"{name} contains NaN or infinity")


def check_graph(data: Data, dataset_index: int) -> Dict[str, Any]:
    required = (
        "x",
        "edge_index",
        "edge_attr",
        "pos",
        "is_protein",
        "y_true",
        "y_pred",
        "y_grt",
    )
    missing = [name for name in required if not hasattr(data, name)]
    if missing:
        raise RuntimeError(f"graph {dataset_index} is missing fields: {missing}")

    num_nodes = int(data.num_nodes)
    if data.x.shape != (num_nodes, 82) or data.x.dtype != torch.float32:
        raise RuntimeError(
            f"graph {dataset_index}: x={tuple(data.x.shape)} {data.x.dtype}"
        )
    if data.edge_index.ndim != 2 or data.edge_index.shape[0] != 2:
        raise RuntimeError(
            f"graph {dataset_index}: edge_index={tuple(data.edge_index.shape)}"
        )
    if data.edge_index.dtype != torch.int64:
        raise RuntimeError(
            f"graph {dataset_index}: edge_index dtype={data.edge_index.dtype}"
        )
    num_edges = int(data.edge_index.shape[1])
    if data.edge_attr.shape != (num_edges, 4):
        raise RuntimeError(
            f"graph {dataset_index}: edge_attr={tuple(data.edge_attr.shape)}"
        )
    if data.edge_attr.dtype != torch.float32:
        raise RuntimeError(
            f"graph {dataset_index}: edge_attr dtype={data.edge_attr.dtype}"
        )

    node_shapes = {
        "pos": (num_nodes, 3),
        "is_protein": (num_nodes, 1),
        "y_true": (num_nodes, 1),
        "y_pred": (num_nodes, 3),
        "y_grt": (num_nodes, 3),
    }
    for name, expected in node_shapes.items():
        tensor = getattr(data, name)
        if tuple(tensor.shape) != expected or tensor.dtype != torch.float32:
            raise RuntimeError(
                f"graph {dataset_index}: {name}={tuple(tensor.shape)} {tensor.dtype}"
            )
        ensure_finite(f"graph {dataset_index} {name}", tensor)

    ensure_finite(f"graph {dataset_index} x", data.x)
    ensure_finite(f"graph {dataset_index} edge_attr", data.edge_attr)
    if not torch.equal(data.pos, data.y_pred):
        raise RuntimeError(f"graph {dataset_index}: pos and y_pred differ")
    if num_edges:
        edge_min = int(data.edge_index.min().item())
        edge_max = int(data.edge_index.max().item())
        if edge_min < 0 or edge_max >= num_nodes:
            raise RuntimeError(
                f"graph {dataset_index}: edge endpoints [{edge_min},{edge_max}] "
                f"outside [0,{num_nodes})"
            )
    protein_values = torch.unique(data.is_protein)
    if not bool(torch.all((protein_values == 0) | (protein_values == 1)).item()):
        raise RuntimeError(f"graph {dataset_index}: is_protein is not binary")

    return {
        "dataset_index": dataset_index,
        "num_nodes": num_nodes,
        "num_edges": num_edges,
        "num_protein_nodes": int(data.is_protein.sum().item()),
        "max_y_true": float(data.y_true.max().item()) if num_nodes else 0.0,
    }


def check_batch(batch: Batch, expected_graphs: int) -> Dict[str, Any]:
    actual_graphs = int(batch.num_graphs)
    if actual_graphs != expected_graphs:
        raise RuntimeError(
            f"batch reports {actual_graphs} graphs, expected {expected_graphs}"
        )
    num_nodes = int(batch.x.shape[0])
    if batch.x.ndim != 2 or batch.x.shape[1] != 82:
        raise RuntimeError(f"batched x has shape {tuple(batch.x.shape)}")
    if batch.edge_attr.ndim != 2 or batch.edge_attr.shape[1] != 4:
        raise RuntimeError(
            f"batched edge_attr has shape {tuple(batch.edge_attr.shape)}"
        )
    if batch.batch.numel() != num_nodes:
        raise RuntimeError("PyG batch assignment length does not match node count")
    if batch.ptr.numel() != actual_graphs + 1:
        raise RuntimeError("PyG batch ptr length is invalid")
    if not torch.equal(batch.pos, batch.y_pred):
        raise RuntimeError("batched pos and y_pred differ")
    ensure_finite("batch x", batch.x)
    ensure_finite("batch edge_attr", batch.edge_attr)
    ensure_finite("batch pos", batch.pos)
    return {
        "num_graphs": actual_graphs,
        "num_nodes": num_nodes,
        "num_edges": int(batch.edge_index.shape[1]),
    }


def evenly_spaced(values: Sequence[int], limit: int) -> List[int]:
    if limit <= 0 or len(values) <= limit:
        return list(values)
    if limit == 1:
        return [values[0]]
    positions = {
        round(index * (len(values) - 1) / (limit - 1)) for index in range(limit)
    }
    return [values[position] for position in sorted(positions)]


def global_indices_by_shard(dataset: CompactGraphDataset) -> List[List[int]]:
    result: List[List[int]] = [[] for _ in dataset.shards]
    graph_map = dataset.manifest.get("graph_map")
    if graph_map is not None:
        for global_index, entry in enumerate(graph_map):
            if isinstance(entry, Mapping):
                shard_index = entry.get("shard", entry.get("shard_index"))
            else:
                shard_index = entry[0]
            result[int(shard_index)].append(global_index)
    else:
        global_index = 0
        for shard_index, count in enumerate(dataset._shard_counts):
            result[shard_index].extend(range(global_index, global_index + count))
            global_index += count
    empty = [index for index, indices in enumerate(result) if not indices]
    if empty:
        raise RuntimeError(f"manifest contains empty shards: {empty}")
    return result


def load_direct(
    dataset: CompactGraphDataset,
    indices: Sequence[int],
) -> Dict[str, Any]:
    before = resource_snapshot()
    latencies: List[float] = []
    samples: List[Dict[str, Any]] = []
    total_nodes = 0
    total_edges = 0
    for dataset_index in indices:
        started = time.perf_counter()
        graph = dataset[dataset_index]
        latency = time.perf_counter() - started
        metrics = check_graph(graph, dataset_index)
        metadata = dataset.metadata(dataset_index)
        metrics.update(
            {
                "shard_index": int(metadata["shard_index"]),
                "source_graph_index": int(metadata["source_graph_index"]),
                "latency_ms": latency * 1000.0,
            }
        )
        if "system_id" in metadata:
            metrics["system_id"] = metadata["system_id"]
        samples.append(metrics)
        latencies.append(latency)
        total_nodes += metrics["num_nodes"]
        total_edges += metrics["num_edges"]
        del graph
    gc.collect()
    after = resource_snapshot()
    resources = resource_delta(before, after)
    elapsed = float(resources["elapsed_s"] or 0.0)
    return {
        "indices": list(indices),
        "latency": latency_summary(latencies),
        "total_nodes": total_nodes,
        "total_edges": total_edges,
        "graphs_per_s": len(indices) / elapsed if elapsed > 0 else None,
        "nodes_per_s": total_nodes / elapsed if elapsed > 0 else None,
        "samples": samples,
        "resources": resources,
    }


def run_loader(
    dataset: CompactGraphDataset,
    indices: Sequence[int],
    *,
    batch_size: int,
    num_workers: int,
    max_batches: int,
    worker_timeout_s: float,
) -> Dict[str, Any]:
    before = resource_snapshot()
    subset = Subset(dataset, list(indices))
    loader = DataLoader(
        subset,
        batch_size=batch_size,
        shuffle=False,
        num_workers=num_workers,
        persistent_workers=False,
        pin_memory=False,
        timeout=worker_timeout_s if num_workers else 0,
    )
    latencies: List[float] = []
    batch_metrics: List[Dict[str, Any]] = []
    iterator = iter(loader)
    prior = time.perf_counter()
    try:
        for batch_number, batch in enumerate(iterator):
            now = time.perf_counter()
            latency = now - prior
            latencies.append(latency)
            metrics = check_batch(batch, min(batch_size, len(indices) - batch_number * batch_size))
            metrics["batch_number"] = batch_number
            metrics["latency_ms"] = latency * 1000.0
            batch_metrics.append(metrics)
            del batch
            if batch_number + 1 >= max_batches:
                break
            prior = time.perf_counter()
    finally:
        del iterator
        del loader
        del subset
    gc.collect()
    after = resource_snapshot()
    resources = resource_delta(before, after)
    elapsed = float(resources["elapsed_s"] or 0.0)
    total_graphs = sum(item["num_graphs"] for item in batch_metrics)
    total_nodes = sum(item["num_nodes"] for item in batch_metrics)
    total_edges = sum(item["num_edges"] for item in batch_metrics)
    return {
        "num_workers": num_workers,
        "worker_timeout_s": worker_timeout_s if num_workers else 0,
        "batch_size": batch_size,
        "input_indices": list(indices),
        "batches_tested": len(batch_metrics),
        "graphs_tested": total_graphs,
        "total_nodes": total_nodes,
        "total_edges": total_edges,
        "graphs_per_s": total_graphs / elapsed if elapsed > 0 else None,
        "nodes_per_s": total_nodes / elapsed if elapsed > 0 else None,
        "latency": latency_summary(latencies),
        "batches": batch_metrics,
        "resources": resources,
    }


def write_report(path: Path, report: Mapping[str, Any]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    temporary = path.with_name(f".{path.name}.tmp.{os.getpid()}")
    with temporary.open("w", encoding="utf-8") as handle:
        json.dump(report, handle, indent=2, sort_keys=True, ensure_ascii=False)
        handle.write("\n")
    os.replace(temporary, path)


def run(args: argparse.Namespace) -> Dict[str, Any]:
    compact = Path(args.compact).expanduser().resolve()
    default_report = compact / (
        f"smoke_report_{os.environ.get('SLURM_JOB_ID', str(os.getpid()))}.json"
    )
    report_path = (
        Path(args.report).expanduser().resolve() if args.report else default_report
    )
    report: Dict[str, Any] = {
        "status": "running",
        "compact": str(compact),
        "report": str(report_path),
        "started_utc_epoch_s": time.time(),
        "environment": {
            "hostname": socket.gethostname(),
            "pid": os.getpid(),
            "python": sys.version,
            "torch": torch.__version__,
            "torch_geometric": torch_geometric.__version__,
            "slurm_job_id": os.environ.get("SLURM_JOB_ID"),
            "slurm_array_job_id": os.environ.get("SLURM_ARRAY_JOB_ID"),
            "slurm_array_task_id": os.environ.get("SLURM_ARRAY_TASK_ID"),
            "slurm_cpus_per_task": os.environ.get("SLURM_CPUS_PER_TASK"),
        },
        "config": vars(args),
        "resources_at_start": resource_snapshot(),
    }
    try:
        torch.set_num_threads(args.torch_threads)
        initialization_before = resource_snapshot()
        dataset = CompactGraphDataset(
            compact,
            max_cached_shards=args.max_cached_shards,
            strict=not args.no_strict,
        )
        initialization_after = resource_snapshot()
        if len(dataset) < 1:
            raise RuntimeError("compact dataset is empty")

        report["dataset"] = {
            "num_graphs": len(dataset),
            "num_shards": len(dataset.shards),
            "n_systems": dataset.manifest.get("n_systems"),
            "method": dataset.manifest.get("method"),
            "size": dataset.manifest.get("size"),
            "initialization": resource_delta(
                initialization_before, initialization_after
            ),
        }

        shard_indices = global_indices_by_shard(dataset)
        selected_shards = evenly_spaced(
            list(range(len(shard_indices))), args.max_shards
        )
        boundary_indices: List[int] = []
        first_per_shard: List[int] = []
        for shard_index in selected_shards:
            indices = shard_indices[shard_index]
            first_per_shard.append(indices[0])
            boundary_indices.append(indices[0])
            if indices[-1] != indices[0]:
                boundary_indices.append(indices[-1])
        report["shard_probe"] = {
            "selected_shards": selected_shards,
            "boundary_indices": boundary_indices,
            "all_shards_selected": len(selected_shards) == len(dataset.shards),
        }
        report["direct_cross_shard"] = load_direct(dataset, boundary_indices)

        rng = random.Random(args.seed)
        random_count = min(args.num_random, len(dataset))
        random_indices = rng.sample(range(len(dataset)), random_count)
        report["direct_random"] = load_direct(dataset, random_indices)

        repeat_index = random_indices[0]
        repeat_latencies: List[float] = []
        repeat_before = resource_snapshot()
        for _ in range(args.repeat_count):
            started = time.perf_counter()
            graph = dataset[repeat_index]
            repeat_latencies.append(time.perf_counter() - started)
            check_graph(graph, repeat_index)
            del graph
        gc.collect()
        repeat_after = resource_snapshot()
        report["repeated_access"] = {
            "index": repeat_index,
            "latency": latency_summary(repeat_latencies),
            "resources": resource_delta(repeat_before, repeat_after),
        }

        cross_loader_indices = first_per_shard[
            : args.batch_size * args.max_batches
        ]
        report["cross_shard_dataloader"] = run_loader(
            dataset,
            cross_loader_indices,
            batch_size=min(args.batch_size, len(cross_loader_indices)),
            num_workers=0,
            max_batches=args.max_batches,
            worker_timeout_s=args.worker_timeout_s,
        )

        worker_indices = list(dict.fromkeys(random_indices + boundary_indices))
        worker_indices = worker_indices[: args.batch_size * args.max_batches]
        # Forked workers should start with an empty mmap cache.  Constructing a
        # fresh dataset here also catches errors in opening the same manifest
        # independently from more than one process.
        worker_dataset = CompactGraphDataset(
            compact,
            max_cached_shards=args.max_cached_shards,
            strict=not args.no_strict,
        )
        report["multiworker_dataloader"] = run_loader(
            worker_dataset,
            worker_indices,
            batch_size=min(args.batch_size, len(worker_indices)),
            num_workers=args.num_workers,
            max_batches=args.max_batches,
            worker_timeout_s=args.worker_timeout_s,
        )
        del worker_dataset

        report["resources_at_end"] = resource_snapshot()
        report["completed_utc_epoch_s"] = time.time()
        report["elapsed_s"] = (
            report["completed_utc_epoch_s"] - report["started_utc_epoch_s"]
        )
        report["status"] = "passed"
    except Exception as error:
        report["status"] = "failed"
        report["completed_utc_epoch_s"] = time.time()
        report["elapsed_s"] = (
            report["completed_utc_epoch_s"] - report["started_utc_epoch_s"]
        )
        report["error"] = f"{type(error).__name__}: {error}"
        report["traceback"] = traceback.format_exc()
        write_report(report_path, report)
        raise

    write_report(report_path, report)
    print(
        json.dumps(
            {
                "status": report["status"],
                "report": str(report_path),
                "num_graphs": report["dataset"]["num_graphs"],
                "num_shards": report["dataset"]["num_shards"],
                "elapsed_s": report["elapsed_s"],
                "rss_mib": report["resources_at_end"]["rss_mib"],
                "hwm_mib": report["resources_at_end"]["hwm_mib"],
            },
            indent=2,
            sort_keys=True,
        ),
        flush=True,
    )
    return report


def main() -> int:
    args = parse_args()
    report = run(args)
    return 0 if report["status"] == "passed" else 1


if __name__ == "__main__":
    raise SystemExit(main())