Virtual-LPU / tests /test_03_bottleneck.py
iparaco's picture
Upload 118 files
e59bd9d verified
Raw History Blame Contribute Delete
3.7 kB
"""Test 03 — Bottleneck classification (mission §9).
Feeds synthetic-but-plausible MEASURED samples into the engine and checks
the decision tree, evidence contents, and honesty labels.
"""
from __future__ import annotations
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from lpu.accelerator.bottleneck import BottleneckEngine
def engine(**over):
cfg = {"bottleneck": {"window_s": 0.0, "min_samples": 3, **over}}
return BottleneckEngine(cfg)
def test_dataloader_bound_from_measured_ratio():
e = engine()
for _ in range(10):
e.feed_session_metrics({"data": 0.8, "forward": 0.1,
"backward": 0.1, "optimizer": 0.02},
data_wait_ratio=0.8)
e.feed_gpu(15.0, 40.0)
rep = e.evaluate(force=True)
assert rep is not None
assert rep.category == "DataLoader Bound"
assert rep.confidence == "Measured"
assert rep.evidence["data_wait_ratio_mean"] == 0.8
assert rep.expected_impact == "Estimate Only"
def test_gpu_compute_bound():
e = engine()
for _ in range(10):
e.feed_gpu(96.0, 60.0)
e.feed_session_metrics({"data": 0.02, "forward": 0.5, "backward": 0.4,
"optimizer": 0.05}, data_wait_ratio=0.02)
rep = e.evaluate(force=True)
assert rep.category == "GPU Compute Bound"
assert rep.evidence["gpu_util_mean_pct"] == 96.0
def test_vram_capacity_bound_blocks_batch_increase():
e = engine()
for _ in range(10):
e.feed_gpu(50.0, 97.0)
rep = e.evaluate(force=True)
assert rep.category == "VRAM Capacity Bound"
assert "Do NOT increase batch" in rep.recommended_action
def test_insufficient_data_is_explicit():
e = engine(min_samples=50)
rep = e.evaluate(force=True)
assert rep is not None
assert rep.category == "Insufficient Data"
assert rep.confidence == "Estimated"
def test_estimated_dataloader_bound_without_sdk():
"""GPU idle + busy CPU without SDK timings => Estimated label, honest note."""
e = engine()
for _ in range(10):
e.feed_gpu(20.0, 30.0)
e.feed_process(350.0)
rep = e.evaluate(force=True)
assert rep.category == "DataLoader Bound (estimated)"
assert rep.confidence == "Estimated"
assert "no SDK phase timings" in rep.evidence["note"]
def test_validation_bound():
e = engine()
for _ in range(5):
e.feed_session_metrics({"data": 0.05, "forward": 0.2,
"backward": 0.1, "optimizer": 0.02,
"validation": 3.0})
rep = e.evaluate(force=True)
assert rep.category == "Validation Bound"
def test_checkpoint_bound():
e = engine()
for _ in range(5):
e.feed_session_metrics({"data": 0.05, "forward": 0.2,
"backward": 0.1, "optimizer": 0.02,
"checkpoint": 1.0})
rep = e.evaluate(force=True)
assert rep.category == "Checkpoint I/O Bound"
def test_history_tracks_category_changes():
e = engine()
for _ in range(6):
e.feed_session_metrics({"data": 0.8, "forward": 0.1}, data_wait_ratio=0.8)
rep1 = e.evaluate(force=True)
assert rep1.category == "DataLoader Bound"
for _ in range(6):
e.feed_session_metrics({"data": 0.02, "forward": 0.5, "backward": 0.4,
"optimizer": 0.05}, data_wait_ratio=0.02)
e.feed_gpu(95.0, 60.0)
rep2 = e.evaluate(force=True)
assert rep2.category == "GPU Compute Bound"
cats = [h["category"] for h in e.history]
assert cats == ["DataLoader Bound", "GPU Compute Bound"] # transitions only