Multi-level Adaptive Execution Tracing for Efficient Performance Analysis

Catalog abstract summary

The available program synopsis describes dynamic kernel/application tracing guided by time-series trends with an emphasis on reducing tracing overhead; the exact trigger logic and evaluation were not captured.

Source: OpenAlex abstract metadata and SERA program/synopsis reviewed; publisher abstract not reproduced because reuse permission was not established.

Problem and motivation

Fixed tracing policies either miss the component responsible for a performance problem or collect excessive event volume and overhead. A useful controller should begin with inexpensive evidence, identify suspicious call or system-call families, and refine tracing around them.

Method and contribution

The workflow has three stages: a pre-test stress signal, perf call-stack sampling at 10 ms intervals, and grouping/ranking of methods and system-call families, retaining the top 5% for deeper tracing. It uses ARIMA(1,1,1) forecasting and an anomaly threshold/score beta to adapt tracepoints, maps system calls to kernel components, and limits call-stack depth to 20 in the stated setup. Firefox is instrumented for cases including Bug 1768481's VsyncRefreshDriverTimer/RunRefreshDrivers path and Bugs 1637586/1776469.

Findings and evidence

Across the reported experiments, approximately 22% of changes were judged useful and 56% were traced; the adaptive configuration reduces coefficient-of-variation overhead by about 95% and ARIMA-related overhead by about 44%, with perf CPU overhead at or below 1% in the stated cases. The Firefox examples show how the controller can move from application-level symptoms to a narrower kernel/user trace scope.

Limitations and future directions

Limitations: Thresholds, beta, the top-5% cutoff, and stack depth are manually selected, and the usefulness/traced-change measures are case-study dependent. The evaluation does not establish robustness under rare faults, distributed systems, changing hardware, or workloads whose symptoms do not appear in the pre-test signal.

Future work: Learn thresholds online, evaluate trace-trigger recall and false negatives, test distributed and containerized workloads, incorporate confidence/abstention, and compare adaptive tracing against fixed minimal/full instrumentation under equal diagnostic budgets.

Sources and identifiers

• Published version published
• SERA program public_source_record

When to cite this paper

Cite this paper when narrowing tracing scope adaptively from lightweight performance evidence to suspicious application/kernel paths.

• The three-stage stress, perf-call-stack, and selective-tracing workflow.
• ARIMA-based adaptation and system-call-to-kernel-component mapping.
• The Firefox case studies and reported overhead reductions.
• The manually tuned threshold, cutoff, and stack-depth limitations.

Citation

@inproceedings{ezzatiJivan2023multilevel,
  author = {Mohammed Adib Khan and Naser Ezzati-Jivan},
  title = {Multi-level Adaptive Execution Tracing for Efficient Performance Analysis},
  year = {2023},
  booktitle = {2023 IEEE International Conference on Software Engineering and Research (SERA)},
  pages = {104-109},
  publisher = {IEEE},
  doi = {10.1109/SERA57763.2023.10197790},
  url = {https://doi.org/10.1109/SERA57763.2023.10197790}
}

Khan, M. A., & Ezzati-Jivan, N. (2023). Multi-level Adaptive Execution Tracing for Efficient Performance Analysis. In 2023 IEEE International Conference on Software Engineering and Research (SERA) (pp. 104-109). https://doi.org/10.1109/SERA57763.2023.10197790

M. A. Khan and N. Ezzati-Jivan, "Multi-level Adaptive Execution Tracing for Efficient Performance Analysis," in 2023 IEEE International Conference on Software Engineering and Research (SERA), pp. 104-109, 2023, doi: 10.1109/SERA57763.2023.10197790