archi829/bottleneck-oracle-data

Bottleneck Oracle Graph Dataset

Model: https://huggingface.co/username/heterogat-v2 License: MIT

---

Overview

The Bottleneck Oracle Dataset is a graph-based benchmark designed for bottleneck detection and performance analysis in computational workloads.

It consists of 501 directed acyclic graphs (DAGs) derived from PyTorch profiler traces of Transformer models, with additional synthetic augmentation to improve structural diversity.

Each graph represents an execution trace where:

• Nodes → computational operations
• Edges → dependency relationships

---

Repository Structure

raw/
 └── traces_backup.zip          # Raw profiler traces

processed/
 └── graphs_v2.pkl             # Training-ready graph dataset

results/
 └── graphrag_results.csv      # Results from GraphRAG optimizer notebook

---

Dataset Structure

Graph Format

• PyTorch Geometric HeteroData
• Directed Acyclic Graphs (DAGs)

Statistics

• Total graphs: 501
• Average nodes per graph: ~22–25
• Minimum nodes: 10 (after filtering)

---

Data Source

Traces are generated using PyTorch profiler on Transformer architectures:

Config	d_model	nhead	num_layers
Tiny	64	2	1
Small	128	4	2
Medium	256	8	4

• Each configuration contributes 167 traces
• Total: 501 graphs

Synthetic DAGs are added when required to maintain dataset diversity.

---

Graph Semantics

• Nodes: Computational operations (e.g., matrix ops, attention blocks)
• Edges: depends_on relationships
• Graphs preserve real execution behavior including parallelism and fork-join patterns

---

Features

Each node contains:

1. Duration (normalized per graph)
2. Start time (normalized)
3. In-degree
4. Out-degree
5. Critical path indicator (binary)
6. Topological rank (normalized)

⚠️ Slack is intentionally excluded to prevent label leakage.

---

Targets

Node-Level Target (y_slack)

• Computed using Critical Path Method (CPM)
• Z-score normalized per graph
• Slack = 0 → node lies on critical path

Graph-Level Target (y_step)

• Total execution time
• Transformed using log₁p + z-score normalization

---

Data Processing

Filtering

• Removed near-chain graphs (critical path ratio > 0.80)
• Removed low-variance slack graphs (std < 0.05)
• Minimum node threshold: 10

Normalization

• Per-graph z-score normalization applied to features and targets

Splits

• Train: ~70%
• Validation: ~15%
• Test: ~15%

---

Pipeline

The dataset is generated using the following notebook pipeline:

nb2 → nb4 → nb3

• nb4 produces the final processed dataset
• graphs_v2.pkl is directly used for training and evaluation

---

Usage

import pickle

with open("processed/graphs_v2.pkl", "rb") as f:
    graphs = pickle.load(f)

---

Notes

• graphs_v2.pkl is the final processed dataset — no intermediate steps required for usage
• Raw traces are preserved for reproducibility
• GraphRAG results are available for downstream evaluation and comparison

---