Skip to content

SMD (Server Machine Dataset)

What is this data? Five weeks of server-machine telemetry from a large internet company: 38 channels per machine (CPU load, network, memory, disk I/O, ...) sampled every minute. Anomalies are real operational incidents annotated by domain experts. The canonical benchmark for multivariate server monitoring.

  • Source: OmniAnomaly (Su et al., KDD 2019) — https://github.com/NetManAIOps/OmniAnomaly
  • Contents: 28 machines (machine-1-1 … 3-11), 38 dimensions, ~25–30k steps each for train and test, 1-minute sampling
  • License: repository is MIT. Only the data files are used.
  • Download: tsad-forge download smd (all machines) or --subset machine-1-1,machine-2-1
  • Loader: load_smd(machine="machine-1-1")

Known flaws

  • Event-length variance: anomaly events range from a few steps to thousands of steps depending on the machine — reporting only machine averages is misleading. This repository stores per-machine results separately.
  • Near-constant channels: some channels are almost constant — zero-variance handling is required during normalization (the loader handles this).
  • Train cleanliness: the claim that the train split is anomaly-free rests on the original authors' assertion.

Quick EDA (from local data)

Sample test channels with labeled anomalies shaded, plus the event-length distribution. Regenerate with tsad-forge viz after downloading the data.