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

Modern system diagnostics rely on kernel-level execution traces to identify bottlenecks and failure modes. However, collecting these traces in production is often prohibited by high runtime overhead, storage limitations, and privacy concerns. The authors ask whether a generative model can produce synthetic kernel traces that are structurally valid and statistically representative enough to substitute for or augment real data in downstream machine learning tasks.

The authors introduce TraceSynth, a framework that models kernel traces as multi-channel sequences—including event types, timestamps, CPU affinity, thread IDs, and process metadata. The core of the system is a Transformer-based denoising diffusion model that learns to generate these sequences from scratch. To ensure the generated data adheres to system-level invariants (such as valid event transitions or CPU affinity), the authors implement a constraint-guided repair module. This module learns rules directly from real traces and performs post-hoc adjustments to the synthetic output, ensuring the generated data is structurally sound.

TraceSynth demonstrates that synthetic data can effectively augment training sets, particularly for deterministic, compute-heavy workloads like scimark2, where it achieves an F1-Macro score of 87.2%—only 2.6 percentage points below models trained on real data. The study identifies context length as the primary driver of quality, with a length of 4096 providing a 104% relative improvement over a length of 256. Furthermore, the authors show that lightweight 2-channel models can retain 97–99% of the performance of full 6-channel models, offering a significant reduction in computational cost for industrial deployment.

This work provides a practical pathway for organizations to perform system diagnostics without the performance penalty of continuous production tracing. By enabling the generation of rare or complex execution scenarios, TraceSynth allows engineers to stress-test their diagnostic pipelines against edge cases that are otherwise difficult to capture, ultimately improving the reliability of system observability tools.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.