TraceSynth generates production-quality kernel traces with diffusion models
Achieves 87.2% F1-Macro, only 2.6 points below real traces at half the cost.
TraceSynth tackles a critical bottleneck in production observability: collecting real kernel execution traces is expensive due to runtime overhead, storage demands, and privacy constraints. The paper, by Yuvraj Sehgal and colleagues from academia and industry, introduces a diffusion model that treats traces as multi-channel sequences (event types, timestamps, CPU affinity, thread IDs, process metadata). A Transformer-based denoising process generates synthetic traces, while constraint-guided repair enforces system invariants to improve realism.
Across six benchmarks, TraceSynth shows strong results for deterministic compute-heavy workloads like scimark2: synthetic augmentation hits 87.2% F1-Macro at context length 4096, only 2.6 percentage points below models trained on real traces alone. Context length is the dominant quality factor—L=4096 yields a +104% relative improvement over L=256. Constraint-guided repair boosts synthetic quality by up to 4.3%. Ablation studies reveal that lightweight 2-channel models (dropping less critical feature channels) retain 97-99% of full 6-channel performance at roughly half the compute cost, making the approach practical for production pipelines.
- TraceSynth uses a Transformer-based diffusion model with constraint-guided repair to generate synthetic kernel traces for ML diagnostics.
- At context length 4096, synthetic augmentation achieves 87.2% F1-Macro, just 2.6 percentage points below real-only baselines for compute-heavy workloads.
- Lightweight 2-channel models retain 97-99% performance of full 6-channel models at half the computational cost.
Why It Matters
Reduces cost and privacy risks of collecting real kernel traces for ML-driven system diagnostics.