TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models

Problem and motivation

Production kernel traces are expensive to collect because of runtime overhead, storage, privacy constraints, and poor coverage of rare behavior. Existing generators do not reliably preserve long-range temporal structure, multi-attribute correlations, and valid execution behavior.

Method and contribution

TraceSynth decodes LTTng kernel traces into Parquet records and overlapping NPZ windows with six channels: event type, inter-event time, CPU, thread ID, command name, and return value. A Transformer denoising diffusion model generates complete traces; learned transition, temporal, CPU-affinity, and attribute constraints detect violations and a post-hoc repair step samples valid values. Downstream utility is measured with next-event prediction on held-out real traces.

Findings and evidence

With a 50/50 real-plus-synthetic training set and context length 4096, scimark2 reaches 87.2% macro-F1 versus 89.8% real-only, while I/O-heavy stream and iozone retain large degradation. Average macro-F1 rises from 30.0% at length 256 to 59.9% at length 4096 (+104% relative). Repair improves 12/15 benchmark-context combinations by 0.3-4.3%, but an unpack-linux/4096 outlier drops by 14.2 points after repair. Two-channel event-plus-time diffusion is within roughly 1-3% of richer models on the tested ablation.

Limitations and future directions

Limitations: The evaluation uses six Phoronix benchmarks, a next-event-prediction proxy, Transformer diffusion, greedy repair, and maximum context 4096 on one H100 node. It does not establish utility for anomaly detection, forecasting, safety-critical RCA, proprietary production behavior, multi-tenant interference, or longer contexts. LTTng, OS, and kernel versions are not reported.

Future work: Validate on proprietary Ciena control-plane traces; add explicit state modeling for I/O nondeterminism; use linear-attention or state-space models for longer contexts; and add differential privacy for cross-organization trace sharing.

Sources and identifiers

• Published version published
• Public ACM PDF · PDF public_full_text

When to cite this paper

Cite this paper when your work uses or compares constraint-guided diffusion generation of complete multi-channel kernel traces, distinct from trace imputation.

• constraint-guided diffusion generation of complete multi-channel kernel traces, distinct from trace imputation.
• the result that temporal context length is a stronger quality driver than adding CPU/thread/command/return channels in the tested downstream task.
• the workload-dependent boundary: near-parity on deterministic compute-heavy scimark2 versus substantial degradation on asynchronous I/O-heavy workloads.
• learned post-hoc transition/temporal/affinity/attribute repair as a low-cost validity safeguard, while preserving the unpack-linux/4096 failure case.

Citation

@inproceedings{ezzatiJivan2026tracesynthgenerating,
  author = {Yuvraj Sehgal and Sneh Patel and Mahsa Panahandeh and Naser Ezzati-Jivan and Francois Tetreault},
  title = {TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models},
  year = {2026},
  booktitle = {ACM International Conference on the Foundations of Software Engineering (FSE) Companion},
  pages = {496-506},
  publisher = {ACM},
  doi = {10.1145/3803437.3805222},
  url = {https://doi.org/10.1145/3803437.3805222}
}

Sehgal, Y., Patel, S., Panahandeh, M., Ezzati-Jivan, N., & Tetreault, F. (2026). TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models. In ACM International Conference on the Foundations of Software Engineering (FSE) Companion (pp. 496-506). https://doi.org/10.1145/3803437.3805222

Y. Sehgal, S. Patel, M. Panahandeh, N. Ezzati-Jivan, and F. Tetreault, "TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models," in ACM International Conference on the Foundations of Software Engineering (FSE) Companion, pp. 496-506, 2026, doi: 10.1145/3803437.3805222