CloudPerfTrace: A High-Resolution Dataset for VM Performance Prediction

CloudPerfTrace is a large-scale dataset featuring 206 system-level metrics captured at a 1-second resolution over 317 days. It is specifically designed for black-box multi-tenant cloud environments where internal VM telemetry is unavailable.

The dataset captures 11 diverse application tasks and the complex interplay between intrinsic workload variations and external resource interference, collected entirely from the host-level hypervisor to respect privacy constraints.

This repository contains a dataset stored in Parquet format and partitioned by application tasks. The dataset is organized under the parquet_ds/ directory, where each partition corresponds to a specific application task ID:

parquet_ds/
 ├── tasks=4/
 ├── tasks=5/
 ├── tasks=6/
 ├── tasks=7/
 ├── tasks=9/
 ├── tasks=10/
 ├── tasks=11/
 ├── tasks=13/
 ├── tasks=14/
 ├── tasks=15/
 └── tasks=16/

Task IDs and Applications

Each task ID represents one application type:

• 4: Data Serving
• 5: Redis
• 6: Web Search
• 7: Graph Analytics
• 9: Data Analytics
• 10: MLPerf
• 11: HBase
• 13: Alluxio
• 14: Minio
• 15: TPC-C
• 16: Flink

Dataset Features & Schema

The dataset provides 206 metrics, split equally between the target VM (_self) and concurrent neighbors (_oth):

• perf_ori: Target performance ratio ($0 < \mathcal{P} \le 1$) representing observed vs. ideal performance.
• tr_self / tr_oth: 53 VM-level metrics (e.g., CPU/Memory utilization) via libvirt.
• lin_self / lin_oth: 38 hardware counters (e.g., LLC misses, cycles) via Linux perf.
• td_self / td_oth: 12 Intel Top-Down analysis metrics.
• workload: Numerical identifier for the workload level.
• tasks: Application task ID (Partition Key).

Temporal Coverage: 317 days of continuous recording at 1-second granularity.

Loading the Dataset

You can easily load the dataset with PyArrow:

import pyarrow.dataset as ds

dataset = ds.dataset("parquet_ds", format="parquet", partitioning="hive")
print(dataset.schema)

Loading the Dataset

The dataset utilizes Hive-style partitioning to allow for high-performance filtering (predicate pushdown). You should use huggingface_hub to download the repository and pyarrow to load the partitioned data.

pip install pyarrow huggingface_hub pandas

Python Implementation

import pyarrow.dataset as ds
from huggingface_hub import snapshot_download

# 1. Download the dataset snapshot to a local cache
repo_path = snapshot_download(
    repo_id="AmirShahbaz/CloudPerfTrace", 
    repo_type="dataset"
)

# 2. Load the partitioned Parquet dataset
# This creates a 'lazy' dataset object that doesn't load everything into RAM at once
dataset = ds.dataset(
    f"{repo_path}/parquet_ds", 
    format="parquet", 
    partitioning="hive"
)

# 3. Efficiently load a specific task (e.g., Task ID 6: Web Search)
# Filtering at the dataset level avoids loading unnecessary files into memory
web_search_df = dataset.to_table(
    filter=ds.field("tasks") == 6
).to_pandas()

print(f"Loaded {len(web_search_df)} rows for Task 6.")
print(web_search_df.head())

Citation

If you use this dataset in your research, please cite it as:

@misc{cloudformer2025,
  title        = {CloudFormer: An Attention-based Performance Prediction for Public Clouds with Unknown Workload},
  author       = {Shahbazinia, Amirhossein and Huang, Darong and Costero, Luis and Atienza, David},
  howpublished = {arXiv preprint arXiv:2509.03394},
  year         = {2025},
  url          = {https://arxiv.org/abs/2509.03394}
}