The-AnkitPatel/ebpf-network-telemetry

eBPF Network Telemetry Queuing Delay Dataset

Overview

This dataset contains 15,000 high-frequency network telemetry observations collected from an experimental Linux testbed using eBPF/XDP hooks. It records microsecond-scale queue dynamics, switch buffer occupancy, ingress throughput, and packet-loss events across virtual network interfaces under synthetic microburst traffic.

The dataset was compiled to evaluate distribution-free uncertainty quantification (specifically inductive conformal prediction) for predicting one-way queuing delay before bufferbloat triggers tail-drop packet loss.

• Author: Ankit Kumar Patel
• Testbed: Distributed Systems & Telemetry Lab, IIT Roorkee
• License: Apache 2.0
• Format: CSV and Parquet

Schema

Each record represents an aggregated telemetry snapshot over a 10 ms sampling window.

Column	Type	Range / Units	Description
timestamp_ms	int64	Unix epoch (ms)	High-precision sampling timestamp
ingress_mbps	float64	50.0 - 980.0 Mbps	Ingress flow bandwidth across the interface
active_flows	int64	5 - 600	Number of concurrent active TCP connections
arrival_kpps	float64	10.0 - 1500.0 kpps	Packet arrival frequency in thousands of packets per second
buffer_util_pct	float64	1.0% - 99.8%	Switch queue buffer occupancy percentage
loss_rate_pct	float64	0.0% - 12.0%	Tail-drop packet drop percentage during buffer saturation
jitter_ms	float64	0.1 - 25.0 ms	Flow inter-arrival time variation
latency_ms	float64	0.8 - 280.0 ms	Target one-way end-to-end queuing and propagation latency

Testbed Configuration

• Probes: Kernel tracepoints (xdp:xdp_redirect, sock:sock_exceed_buf_limit) and eBPF ring buffers exporting 6-feature vectors.
• Traffic Generation: Multi-flow synthetic packet generator producing heavy-tailed Pareto bursts to simulate datacenter traffic surges.
• Environment: Ubuntu 22.04 LTS, Linux kernel 6.5, simulated multi-tenant veth topology with netem queuing emulation.

Usage

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("The-AnkitPatel/ebpf-network-telemetry", split="train")
df = dataset.to_pandas()
print(df.head())

Pandas

import pandas as pd

url = "https://huggingface.co/datasets/The-AnkitPatel/ebpf-network-telemetry/raw/main/network_telemetry_ebpf.csv"
df = pd.read_csv(url)
print(f"Loaded {len(df)} samples across columns: {list(df.columns)}")

Intended Use

• Evaluation of Split Conformal Prediction, Mondrian Conformal Prediction, and Adaptive Conformal Inference (ACI) under distribution drift.
• Benchmark comparisons between gradient-boosted decision trees (LightGBM) and neural network baselines on tabular systems telemetry.
• Active Queue Management (AQM) and congestion control research.

Related Links

• Model: The-AnkitPatel/netpredict-telemetry-lgbm
• Live Space: The-AnkitPatel/netpredict-conformal-telemetry
• Kaggle Notebook: theankitkumarpatel/conformalized-lightgbm-network-telemetry

Citation

@misc{patel2026ebpftelemetry,
  author = {Patel, Ankit Kumar},
  title = {eBPF Network Telemetry Queuing Delay Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/The-AnkitPatel/ebpf-network-telemetry}}
}