{ "schema_version": "0.6.0", "paper_id": "system-execution-path-profiling-hardware-counters", "page_url": "https://threadslab.org/research-publications/papers/system-execution-path-profiling-hardware-counters/", "title": "System Execution Path Profiling Using Hardware Performance Counters", "title_variants": [], "authors": [ "Francis Giraldeau", "Naser Ezzati-Jivan", "Michel R. Dagenais" ], "author_details": [ { "name": "Francis Giraldeau", "orcid": null, "profile_url": "https://dblp.org/pid/183/1930.html" }, { "name": "Naser Ezzati-Jivan", "orcid": "https://orcid.org/0000-0003-1435-6297", "profile_url": "https://threadslab.org/" }, { "name": "Michel R. Dagenais", "orcid": null, "profile_url": "https://dblp.org/pid/60/309.html" } ], "publication": { "year": 2021, "venue": "2021 IEEE International Systems Conference (SysCon)", "type": "conference paper", "publication_date": "2021-04-15", "online_date": null, "print_date": "2021-04-15", "volume": null, "issue": null, "pages": "1-8", "article_number": null, "publisher": "IEEE", "issn": [], "isbn": [], "crossref_type": "proceedings-article" }, "publication_type": "conference paper", "status": "published_with_related_public_full_text", "canonical_source_url": "https://doi.org/10.1109/SYSCON48628.2021.9447121", "source_record_id": "system-execution-path-profiling-using-hardware-performance-counters-c4cbc41744", "identifiers": { "doi": "10.1109/SYSCON48628.2021.9447121" }, "abstract": "The related public chapter describes low-intrusion execution-path profiling that forwards hardware performance-counter overflows from the kernel to user space, where a CPython profiler records interpreted-code call stacks and LTTng-UST stores the samples.", "abstract_source": "Exact Overleaf-generated author-manuscript candidate cross-checked with public thesis Chapter 6 and the official SysCon program; the thesis is the public full-text route and is not the SysCon Version of Record.", "abstract_available": true, "scholar_eligibility": { "eligible": false, "basis": "not-eligible", "note": "The page is a discovery record; it does not claim Google Scholar article-host eligibility." }, "description": "The related public chapter describes low-intrusion execution-path profiling that forwards hardware performance-counter overflows from the kernel to user space, where a CPython profiler records interpreted-code call stacks and LTTng-UST stores the samples.", "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": "2026-08-28", "sources": [ { "note": "DOI: https://doi.org/10.1109/SYSCON48628.2021.9447121" }, { "note": "DBLP record: https://dblp.org/rec/conf/syscon/GiraldeauJD21" }, { "note": "Official SysCon program: https://events-siteplex.confcats.io/syscon2021/wp-content/uploads/sites/19/2021/04/syscon21-program.pdf" }, { "note": "Public thesis PDF, Chapter 6 pp. 93--110: https://core.ac.uk/download/pdf/213619806.pdf" }, { "note": "Public thesis pp. 93--96: problem, PMU/perf-NMI architecture, PyPMU contributions, and interpreter boundary" }, { "note": "Public thesis pp. 101--106: accuracy, event-forwarding cost, stack unwinding, LTTng-UST comparison, and profiling overhead" }, { "note": "Public thesis pp. 109--110: Trace Compass integration, user-level interrupt direction, signal limitations, and conclusion" }, { "note": "Overleaf project: https://www.overleaf.com/project/5e01a488862790000145cf86" }, { "note": "Original Overleaf PDF: 8 pages; exact title/authors/abstract matched, but an extraneous compile-log fragment was embedded in the body" }, { "note": "Separate cleanup Overleaf copy: source fragment replaced, IEEE author-manuscript notice added, recompiled successfully, and cleaned 8-page PDF inspected on first and final pages" }, { "note": "Rights gate: agreement-specific SysCon self-archiving terms still require confirmation before arXiv submission; do not use the IEEE Version of Record" }, { "note": "Private batch report: reports/parallel-batch-03.md" } ] }, "summary": { "core_contribution": "The work bridges kernel-level hardware-counter sampling and user-space interpreter state so execution-path profiling can attribute performance evidence to interpreted code.", "problem": "Kernel execution paths expose where distributed or heterogeneous tasks wait, but hardware-counter samples normally identify native instructions and lose the connection to interpreted-language statements; profiling inside the kernel also risks violating runtime isolation and NMI-time constraints.", "method": "The related public chapter uses Linux perf to receive PMU overflow notifications in the kernel NMI handler, forwards an event to user space with a signal, and implements PyPMU for CPython. User-space code records interpreter/call-stack state; LTTng-UST writes the samples to a ring buffer, while online and offline libunwind-based stack recovery are compared. The chapter also relates the profiler to kernel-trace execution paths and reports an evaluation on an Intel i7-4770 system with 16 GB RAM, Ubuntu 14.04, and Linux 3.13.", "findings": "The related chapter reports 0.1% RMS profile error on its calibrated function benchmark, 2.93 microseconds average event-forwarding cost over 10^6 samples, and 4.15 microseconds for counter management and signal setup. It reports online unwinding as preferable under the tested depth limit, LTTng-UST as at least 25 times faster than built-in Python traceback recording, and approximately 25% lower trace storage in that comparison. These measurements come from the public thesis chapter and are not claimed as independently rechecked against the IEEE SysCon PDF.", "limitations": "The original Overleaf project contained an extraneous compile-log fragment; a separate cleanup copy removed it, recompiled successfully, and added the IEEE author-manuscript notice. The signed SysCon publication agreement and any venue-specific self-archiving wording were not independently inspected. The public full-text route remains a related 2015 thesis chapter rather than the standalone IEEE SysCon manuscript. The evaluation is centered on CPython, one Linux environment, synthetic/calibrated profiling workloads, and the reported monitoring comparisons; broader runtimes, hardware generations, and production distributed workloads remain unverified.", "future_work": "The chapter identifies integration with task-execution-path extraction and Trace Compass, support for additional runtimes, replacing signals with user-level interrupts, handling signal-blocking limitations, and reducing redundant call-stack trace data as future directions." }, "tags": [ "performance-analysis", "resource-analysis", "multicore-systems", "system-tracing" ], "keywords": [ "execution paths", "hardware performance counters", "critical-path profiling", "CPython", "PyPMU", "Linux perf", "perf NMI", "LTTng-UST", "libunwind", "sampling", "SysCon 2021" ], "versions": [ { "id": "published-version", "label": "Published version", "relation": "version-of-record", "title": "System Execution Path Profiling Using Hardware Performance Counters", "url": "https://doi.org/10.1109/SYSCON48628.2021.9447121", "pdf_url": null, "status": "published", "canonical_for_citation": true }, { "id": "related-public-thesis-chapter-plus-cleaned-local-overleaf-manuscript-candidate-with-ieee-rights-notice", "label": "Related public thesis chapter plus cleaned local Overleaf manuscript candidate with IEEE rights notice", "relation": "source-record", "title": "System Execution Path Profiling Using Hardware Performance Counters", "url": "https://core.ac.uk/download/pdf/213619806.pdf", "pdf_url": "https://core.ac.uk/download/pdf/213619806.pdf", "status": "public_related_full_text", "canonical_for_citation": false } ], "access": { "status": "published_with_related_public_full_text", "note": "The SysCon DOI is the canonical citation target. A public 2015 Polytechnique thesis mirror contains Chapter 6, pp. 93--110, with the exact title and related PMU-to-user-space/CPython/LTTng-UST method. The exact Overleaf source was copied to a separate cleanup project, the compile-log fragment was removed, an IEEE personal-use/author-manuscript notice was added, and an 8-page PDF was recompiled and page-inspected; the final copy contains the DOI and notice without a publisher production footer. It remains an arXiv candidate pending confirmation of the signed SysCon self-archiving terms. The public thesis chapter is not the IEEE Version of Record or a standalone SysCon preprint, and no publisher PDF is redistributed.", "license": null }, "resources": { "code": null, "data": null, "slides": null, "demo": null }, "citation_guidance": { "when_to_cite": "Cite this paper when its specific method, evidence, or benchmark is directly relevant.", "points": [ "The paper's method is directly relevant.", "The paper's evidence or benchmark is directly relevant." ], "canonical_version_id": "published-version" }, "provenance": { "metadata_verified_on": "2026-08-28", "metadata_source": [ "DOI: https://doi.org/10.1109/SYSCON48628.2021.9447121", "DBLP record: https://dblp.org/rec/conf/syscon/GiraldeauJD21", "Official SysCon program: https://events-siteplex.confcats.io/syscon2021/wp-content/uploads/sites/19/2021/04/syscon21-program.pdf", "Public thesis PDF, Chapter 6 pp. 93--110: https://core.ac.uk/download/pdf/213619806.pdf", "Public thesis pp. 93--96: problem, PMU/perf-NMI architecture, PyPMU contributions, and interpreter boundary", "Public thesis pp. 101--106: accuracy, event-forwarding cost, stack unwinding, LTTng-UST comparison, and profiling overhead", "Public thesis pp. 109--110: Trace Compass integration, user-level interrupt direction, signal limitations, and conclusion", "Overleaf project: https://www.overleaf.com/project/5e01a488862790000145cf86", "Original Overleaf PDF: 8 pages; exact title/authors/abstract matched, but an extraneous compile-log fragment was embedded in the body", "Separate cleanup Overleaf copy: source fragment replaced, IEEE author-manuscript notice added, recompiled successfully, and cleaned 8-page PDF inspected on first and final pages", "Rights gate: agreement-specific SysCon self-archiving terms still require confirmation before arXiv submission; do not use the IEEE Version of Record", "Private batch report: reports/parallel-batch-03.md" ], "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": "2026-08-28" } }