{
  "schema_version": "0.6.0",
  "paper_id": "tracesynth-kernel-trace-generation",
  "page_url": "https://naser.github.io/research-publications/papers/tracesynth-kernel-trace-generation/",
  "title": "TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models",
  "title_variants": [],
  "authors": [
    "Yuvraj Sehgal",
    "Sneh Patel",
    "Mahsa Panahandeh",
    "Naser Ezzati-Jivan",
    "Francois Tetreault"
  ],
  "author_details": [
    {
      "name": "Yuvraj Sehgal",
      "orcid": "https://orcid.org/0009-0002-1980-604X",
      "profile_url": "https://dblp.org/pid/434/8410.html"
    },
    {
      "name": "Sneh Patel",
      "orcid": "https://orcid.org/0000-0002-1985-8715",
      "profile_url": "https://dblp.org/pid/315/9849.html"
    },
    {
      "name": "Mahsa Panahandeh",
      "orcid": "https://orcid.org/0000-0002-6369-8982",
      "profile_url": "https://dblp.org/pid/201/1655.html"
    },
    {
      "name": "Naser Ezzati-Jivan",
      "orcid": "https://orcid.org/0000-0003-1435-6297",
      "profile_url": "https://naser.github.io/"
    },
    {
      "name": "Francois Tetreault",
      "orcid": "https://orcid.org/0000-0003-1975-1501",
      "profile_url": "https://dblp.org/pid/288/0256.html"
    }
  ],
  "publication": {
    "year": 2026,
    "venue": "ACM International Conference on the Foundations of Software Engineering (FSE) Companion",
    "type": "conference paper",
    "publication_date": "2026-07-05",
    "online_date": "2026-07-17",
    "print_date": "2026-07-05",
    "volume": null,
    "issue": null,
    "pages": "496-506",
    "article_number": null,
    "publisher": "ACM",
    "issn": [],
    "isbn": [],
    "crossref_type": "proceedings-article"
  },
  "publication_type": "conference paper",
  "status": "published_with_public_full_text",
  "canonical_source_url": "https://doi.org/10.1145/3803437.3805222",
  "source_record_id": "tracesynth-generating-production-quality-kernel-traces-with-constraint-guided-diffusion-models-e5ad60d639",
  "identifiers": {
    "doi": "10.1145/3803437.3805222"
  },
  "abstract": null,
  "abstract_source": "Abstract not reproduced; the four-part summary is paraphrased from the verified full text.",
  "abstract_available": false,
  "scholar_eligibility": {
    "eligible": false,
    "basis": "not-eligible",
    "note": "The page is a discovery record; it does not claim Google Scholar article-host eligibility."
  },
  "description": "TraceSynth generates novel structured kernel-trace windows with a Transformer diffusion model and repairs generated events against invariants mined from real LTTng traces.",
  "evidence_level": "full-text-reviewed",
  "evidence": {
    "source_basis": "full-text-reviewed",
    "coverage": "material paper sections",
    "summary_origin": "AI-assisted catalog editorial summary",
    "review_status": "catalog-reviewed; paper-author approval pending",
    "verified_on": "2026-08-09",
    "sources": [
      {
        "note": "Private full PDF read: pdf-evidence/notes/tracesynth-kernel-trace-generation.md"
      },
      {
        "note": "TraceSynth PDF pp. 3-5: LTTng collection, six-channel representation, diffusion model, and invariant repair"
      },
      {
        "note": "TraceSynth PDF pp. 6-10: Nibicluster/Phoronix workloads, DDIM generation, next-event metrics, context-length and channel ablations"
      },
      {
        "note": "TraceSynth PDF pp. 10-11: limitations and future work"
      },
      {
        "note": "Local PDF hash verified in pdf-evidence/extraction-manifest.json"
      }
    ]
  },
  "summary": {
    "core_contribution": "TraceSynth generates novel structured kernel-trace windows with a Transformer diffusion model and repairs generated events against invariants mined from real LTTng traces.",
    "problem": "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": "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": "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": "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."
  },
  "tags": [
    "kernel-tracing",
    "system-tracing",
    "trace-analysis",
    "machine-learning"
  ],
  "keywords": [
    "kernel traces",
    "trace generation",
    "diffusion models",
    "constraint-guided generation",
    "LTTng",
    "Parquet",
    "Phoronix Test Suite",
    "DDIM",
    "next-event prediction",
    "synthetic traces"
  ],
  "versions": [
    {
      "id": "published-version",
      "label": "Published version",
      "relation": "version-of-record",
      "title": "TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models",
      "url": "https://doi.org/10.1145/3803437.3805222",
      "pdf_url": null,
      "status": "published",
      "canonical_for_citation": true
    },
    {
      "id": "public-acm-pdf",
      "label": "Public ACM PDF",
      "relation": "source-record",
      "title": "TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models",
      "url": "https://dl.acm.org/doi/pdf/10.1145/3803437.3805222",
      "pdf_url": "https://dl.acm.org/doi/pdf/10.1145/3803437.3805222",
      "status": "public_full_text",
      "canonical_for_citation": false
    }
  ],
  "access": {
    "status": "published_with_public_full_text",
    "note": "The DOI is the canonical citation target. The external source PDF is linked for readers; no PDF is redistributed in this repository.",
    "license": null
  },
  "resources": {
    "code": null,
    "data": null,
    "slides": null,
    "demo": null
  },
  "citation_guidance": {
    "when_to_cite": "Cite this paper when your work uses or compares constraint-guided diffusion generation of complete multi-channel kernel traces, distinct from trace imputation.",
    "points": [
      "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."
    ],
    "canonical_version_id": "published-version"
  },
  "provenance": {
    "metadata_verified_on": "2026-08-09",
    "metadata_source": [
      "Private full PDF read: pdf-evidence/notes/tracesynth-kernel-trace-generation.md",
      "TraceSynth PDF pp. 3-5: LTTng collection, six-channel representation, diffusion model, and invariant repair",
      "TraceSynth PDF pp. 6-10: Nibicluster/Phoronix workloads, DDIM generation, next-event metrics, context-length and channel ablations",
      "TraceSynth PDF pp. 10-11: limitations and future work",
      "Local PDF hash verified in pdf-evidence/extraction-manifest.json"
    ],
    "summary_written_by": "AI-assisted",
    "summary_verified_by": "full-text-grounded catalog review; author approval pending",
    "linked_preprint_record": null,
    "author_order_note": null
  },
  "batch": {
    "phase": 2,
    "batch_label": "expanded forty-paper release",
    "status": "included_in_expanded_catalog",
    "selected_at": "2026-08-09"
  }
}