2017 · IEEE High Performance Extreme Computing Conference (HPEC)
Dynamic trace-based sampling algorithm for memory usage tracking of enterprise applications
Evidence basis: full-text-reviewed · Review status: catalog-reviewed; paper-author approval pending
kernel-tracing resource-analysis performance-analysis trace-reduction
memory usage dynamic sampling enterprise applications trace overhead LTTng kernel module Kprobes kmem_mm_page_alloc kmem_mm_page_free
Core contribution: The paper introduces a kernel-resident dynamic sampler that tracks virtual and physical memory usage while reducing the event volume generated by high-frequency memory activity.
Problem and motivation
Full memory-event tracing can generate excessive event volume and perturb enterprise applications, while coarse periodic sampling can miss important memory changes.
Method and contribution
A Linux kernel module dynamically samples virtual and physical memory using a two-dimensional timer/variability policy. It observes allocation-related state and physical-page allocation/free events, maintains concurrent state with an RCU hash map and atomic values, and exposes configuration through /proc.
Findings and evidence
With a 10 ms/10 MB policy, dynamic sampling tracks the Massif memory curves for the Firefox, Totem, and synthetic workloads while emitting far fewer events than all-event tracing. The authors report substantially lower runtime overhead than Massif and large event-volume reductions.
Limitations and future directions
Limitations: The experiments use a synthetic buffer workload plus Firefox and a 10-second Totem video, with no numeric memory-error bound reported. Only one sampling configuration is emphasized, and the paper's prose claim of at most 1% overhead is not fully consistent with rounded table values.
Future work: Extend the dynamic sampler to network usage, disk I/O, and other high-frequency kernel events.
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 a timer-plus-memory-variability policy that reduces memory-trace event volume without the full cost of Massif.
- For a timer-plus-memory-variability policy that reduces memory-trace event volume without the full cost of Massif.
- For the 5 GB/4 GB event-count reductions and the use of kmem page-allocation events as the all-event baseline.
- For a practical LTTng/Linux-kernel implementation of dynamic memory tracing with /proc configuration.
Citation
@inproceedings{ezzatiJivan2017dynamictrace,
author = {Houssem Daoud and Naser Ezzati-Jivan and Michel R. Dagenais},
title = {Dynamic trace-based sampling algorithm for memory usage tracking of enterprise applications},
year = {2017},
booktitle = {IEEE High Performance Extreme Computing Conference (HPEC)},
pages = {1-7},
publisher = {IEEE},
doi = {10.1109/hpec.2017.8091061},
url = {https://doi.org/10.1109/hpec.2017.8091061}
}Other citation formats for Word and reference managers
Daoud, H., Ezzati-Jivan, N., & Dagenais, M. R. (2017). Dynamic trace-based sampling algorithm for memory usage tracking of enterprise applications. In IEEE High Performance Extreme Computing Conference (HPEC) (pp. 1-7). https://doi.org/10.1109/hpec.2017.8091061H. Daoud, N. Ezzati-Jivan, and M. R. Dagenais, "Dynamic trace-based sampling algorithm for memory usage tracking of enterprise applications," in IEEE High Performance Extreme Computing Conference (HPEC), pp. 1-7, 2017, doi: 10.1109/hpec.2017.8091061