File size: 5,183 Bytes
92ff720
 
 
 
 
 
 
 
 
 
 
 
 
 
911259f
 
 
92ff720
911259f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c717a65
911259f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1f8bf6c
911259f
a9fceb6
 
1f8bf6c
a9fceb6
 
 
911259f
 
 
 
 
 
 
 
 
 
 
 
 
 
1f8bf6c
911259f
 
 
 
 
 
1f8bf6c
 
911259f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
731dd74
 
 
 
 
 
 
 
911259f
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
---
license: cc-by-4.0
language:
- en
tags:
- hardware
- infrastructure
- system
- subsystem
- CPU
- GPU
- memory
- network
- storage
- telemetry
- anomaly-detection
- performance
pretty_name: Reveal
---

# 🛰️ Dataset Card for **Reveal: Hardware Telemetry Dataset for Machine Learning Infrastructure Profiling and Anomaly Detection**

## Dataset Details

### Dataset Description

**Reveal** is a large-scale, curated dataset of **hardware telemetry** collected from high-performance computing (HPC) while running diverse machine learning (ML) workloads.  
It enables reproducible research on **system-level profiling**, **unsupervised anomaly detection**, and **ML infrastructure optimization**.  

The dataset accompanies the paper  
📄 *“Detecting Anomalies in Systems for AI Using Hardware Telemetry”* (Chen *et al.*, University of Oxford, 2025).  
Reveal captures low-level hardware and operating system metrics—fully accessible to operators—allowing anomaly detection **without requiring workload knowledge or instrumentation**.

- **Curated by:** Ziji Chen, Steven W. D. Chien, Peng Qian, Noa Zilberman (University of Oxford, Department of Engineering Science)  
- **Shared by:** Ziji Chen (contact: [email protected])  
- **Language(s):** English (metadata and documentation)  
- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

---

### Dataset Sources

- **Paper:** [Detecting Anomalies in Systems for AI Using Hardware Telemetry](https://arxiv.org/abs/2510.26008)  
- **DOI:** [10.5281/zenodo.17470313](https://doi.org/10.5281/zenodo.17470313)  

---

## Uses

### Direct Use

Reveal can be used for:
- Research on **unsupervised anomaly detection** in system telemetry  
- Modeling **multivariate time-series** from hardware metrics  
- Studying **cross-subsystem interactions** (CPU, GPU, memory, network, storage)  
- Developing **performance-aware ML infrastructure tools**  
- Training or benchmarking anomaly detection models for **AIOps** and **ML system health monitoring**

### Out-of-Scope Use

The dataset **should not** be used for:
- Inferring or reconstructing user workloads or model behavior  
- Benchmarking end-user application performance  
- Any use involving personal, confidential, or proprietary data reconstruction  

---

## Dataset Structure

Reveal consists of time-series telemetry, derived features, and automatically labeled anomaly segments.



**Core fields include:**
- `timestamp`: UTC time of sample  
- `host_id`: host or node identifier  
- `metric_name`: name of the measured counter  
- `value`: recorded numeric value  
- `subsystem`: {CPU, *GPU (if supported by the underlying infrastructure), Memory, Network, Storage}  

**Additional Notes**

A complete list of metrics and their descriptions can be found in `MetricDescriptionCPU.md` and `MetricDescriptionGPU.md`.

After downloading and extracting the dataset zip, place the `meta.csv` file and the `example Jupyter notebooks` inside the `Reveal/` directory before running.

---

## Dataset Creation

### Curation Rationale

Modern ML workloads are complex and opaque to operators due to virtualization and containerization. Reveal was created to **enable infrastructure-level observability** and anomaly detection purely from hardware telemetry, without access to user workloads.

### Source Data

#### Data Collection and Processing

- Collected using: `perf`, `procfs`, `nvidia-smi`, and standard Linux utilities  
- Sampling interval: 100 ms  
- ~150 raw metric types per host, expanded to ~700 time-series channels, including metrics related to GPUs. 

#### Workloads and Systems

- **Workloads:** >30 ML applications (BERT, BART, ResNet, ViT, VGG, DeepSeek, LLaMA, Mistral)  
- **Datasets:** GLUE/SST2, WikiSQL, PASCAL VOC, CIFAR, MNIST  
- **Systems:**  
  - Dual-node GPU HPC cluster: Two nodes, each with two NVIDIA V100 GPUs (32 GB), an Intel Xeon Platinum 8628 CPU (48 cores), 384 GB memory, connected through InfiniBand HDR100. Packaged as `Reveal.zip`.
  - Nine-node CPU cluster: Nine servers, each running 11 Apptainer containers (four threads and 20 GB memory per container), powered by AMD EPYC 7443P CPUs. Packaged as `RevealCPURun<n>.zip`.

#### Who are the data producers?

All data was generated by the authors in controlled environments using synthetic workloads.  
No user or private information is included.

### Annotations

#### Personal and Sensitive Information
No personal, identifiable, or proprietary data.  
All records are machine telemetry and anonymized.

---

## Bias, Risks, and Limitations

- Collected on specific hardware (NVIDIA/AMD CPUs, NVIDIA GPUs); behavior may differ on other architectures.  
- Reflects **controlled test conditions**, not production cloud variability.  

---

## Citation

**BibTeX:**
```bibtex
@misc{chen2025detectinganomaliesmachinelearning,
      title={Detecting Anomalies in Machine Learning Infrastructure via Hardware Telemetry}, 
      author={Ziji Chen and Steven W. D. Chien and Peng Qian and Noa Zilberman},
      year={2025},
      eprint={2510.26008},
      archivePrefix={arXiv},
      primaryClass={cs.PF},
      url={https://arxiv.org/abs/2510.26008}, 
}