2018 · IEEE Systems Conference
High latency cause detection using multilevel dynamic analysis
Evidence basis: full-text-reviewed · Review status: catalog-reviewed; paper-author approval pending
system-tracing latency-analysis root-cause-analysis performance-analysis
high latency dynamic analysis multilevel analysis latency causes LTTng LTTng-UST Trace Compass OPcache contention
Core contribution: The paper correlates PHP user-space events with LTTng kernel events in a unified multilevel model to explain high-latency web requests.
Problem and motivation
Web-request latency is caused by interactions among application code, runtime compilation/cache behavior, synchronization, scheduling, and I/O; a user-space trace alone cannot identify all causes.
Method and contribution
Instrument PHP and correlate its call/compile events with LTTng-UST and kernel states in a common timestamped model. Mapping rules and an FSM reconstruct PHP call-stack state; synchronized views and a Critical Path View expose blocking and preemption.
Findings and evidence
The simple example identifies compile/cache time, while the Drupal example identifies OPcache shared-memory lock contention that is invisible in PHP-only analysis. Minimal cross-layer tracing limits overhead to below 7.5% in the reported client sweep.
Limitations and future directions
Limitations: The implementation and evaluation are PHP-specific; the paper does not establish transferability to other runtimes or web stacks. Full tracing has substantial throughput cost, and the compile percentage is case-specific.
Future work: Add mapping rules and views, and investigate machine-learning-based predictive analytics, as stated by the authors.
Sources and identifiers
- Published version published
- Public full text · PDF public_full_text
When to cite this paper
Cite this paper when your work uses or compares multilevel PHP/LTTng/kernel correlation that distinguishes application execution, compilation/cache, scheduling, and blocking states.
- For multilevel PHP/LTTng/kernel correlation that distinguishes application execution, compilation/cache, scheduling, and blocking states.
- For the OPcache shared-memory lock-contention diagnosis in a Drupal/ApacheBench workload.
- For the measured overhead tradeoff between full tracing and a minimal syscall/process-scheduling configuration.
Citation
@inproceedings{ezzatiJivan2018highlatency,
author = {Naser Ezzati-Jivan and Genevieve Bastien and Michel R. Dagenais},
title = {High latency cause detection using multilevel dynamic analysis},
year = {2018},
booktitle = {IEEE Systems Conference},
pages = {1-8},
publisher = {IEEE},
doi = {10.1109/syscon.2018.8369613},
url = {https://doi.org/10.1109/syscon.2018.8369613}
}Other citation formats for Word and reference managers
Ezzati-Jivan, N., Bastien, G., & Dagenais, M. R. (2018). High latency cause detection using multilevel dynamic analysis. In IEEE Systems Conference (pp. 1-8). https://doi.org/10.1109/syscon.2018.8369613N. Ezzati-Jivan, G. Bastien, and M. R. Dagenais, "High latency cause detection using multilevel dynamic analysis," in IEEE Systems Conference, pp. 1-8, 2018, doi: 10.1109/syscon.2018.8369613