Delete Description.md
Browse files- Description.md +0 -43
Description.md
DELETED
|
@@ -1,43 +0,0 @@
|
|
| 1 |
-
# 🛰️ Reveal: Curated Telemetry Dataset for Machine Learning Infrastructure Profiling and Anomaly Detection
|
| 2 |
-
|
| 3 |
-
**Authors:** Ziji Chen, Steven W. D. Chien, Peng Qian, Noa Zilberman
|
| 4 |
-
**Institution:** University of Oxford
|
| 5 |
-
**Paper:** [Detecting Anomalies in Systems for AI Using Hardware Telemetry (arXiv, 2025)](https://arxiv.org/abs/submit/6934461)
|
| 6 |
-
**License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
|
| 7 |
-
**Version:** 1.0
|
| 8 |
-
|
| 9 |
-
---
|
| 10 |
-
|
| 11 |
-
## 📘 Overview
|
| 12 |
-
|
| 13 |
-
**Reveal** is a large-scale, curated telemetry dataset for studying performance profiling, anomaly detection, and infrastructure optimization in modern machine learning (ML) systems.
|
| 14 |
-
It captures **hardware-level signals** from both GPU-equipped clusters while running **over 30 popular ML workloads** across NLP and computer vision domains.
|
| 15 |
-
|
| 16 |
-
This dataset was collected using the **Reveal** framework, a hardware-centric profiling and unsupervised anomaly detection system introduced in our paper.
|
| 17 |
-
Reveal observes **CPU, GPU, memory, network, and storage** metrics without requiring access to user workloads, making the dataset ideal for operator, side anomaly detection and system performance analysis.
|
| 18 |
-
|
| 19 |
-
---
|
| 20 |
-
|
| 21 |
-
## 🧠 Motivation
|
| 22 |
-
|
| 23 |
-
Modern ML systems are tightly coupled across hardware and software layers, yet operators often lack visibility into workloads due to virtualization and containerization.
|
| 24 |
-
Reveal bridges this gap by providing a **hardware-only telemetry view**, enabling anomaly detection and performance diagnosis **without application instrumentation**.
|
| 25 |
-
|
| 26 |
-
---
|
| 27 |
-
|
| 28 |
-
## 🧩 Dataset Description
|
| 29 |
-
|
| 30 |
-
| Category | Description |
|
| 31 |
-
|-----------|-------------|
|
| 32 |
-
| **Sampling rate** | 100 ms per metric |
|
| 33 |
-
| **Metric types** | ~150 raw metric types per host |
|
| 34 |
-
| **Subsystems covered** | CPU, GPU, Memory, Network, Disk |
|
| 35 |
-
| **Time-series channels** | ~700 per host |
|
| 36 |
-
| **Workloads** | 30+ ML applications including BERT, BART, ResNet, ViT, VGG, DeepSeek, LLaMA, Mistral |
|
| 37 |
-
| **Datasets used** | GLUE/SST2, WikiSQL, PASCAL VOC, CIFAR, MNIST |
|
| 38 |
-
| **Systems** | Dual-host GPU HPC cluster |
|
| 39 |
-
| **Telemetry tools** | `perf`, `procfs`, `nvidia-smi`, Linux utilities |
|
| 40 |
-
|
| 41 |
-
Each record corresponds to a **time-series window** of low-level system metrics.
|
| 42 |
-
|
| 43 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|