2021 · IEEE International Conference on Cloud Computing Technology and Science (CloudCom)
Container Workload Characterization Through Host System Tracing
Evidence basis: metadata-only · Review status: catalog-reviewed; paper-author approval pending
kernel-tracing system-tracing performance-analysis resource-analysis
container workloads host tracing LTTng pid_ns tid Trace Compass PageRank K-Means Docker Ubuntu 20.04.1 Linux 5.8.0 silhouette score tracing overhead
Core contribution: The paper characterizes container workloads from host-level LTTng traces by selecting important threads and clustering execution-state behavior without instrumenting the containers internally.
Problem and motivation
Cloud users may not control internal container agents, and coarse CPU, disk, and network metrics do not reveal enough execution detail for workload characterization.
Method and contribution
LTTng kernel tracing records host events with pid_ns and tid context so events can be attributed to containers. Per-thread states such as user mode, system-call mode, preempted, interrupted, blocked, and waiting are aggregated into normalized container vectors; an adapted PageRank selects important threads and two-stage K-Means produces coarse and fine workload groups. Trace Compass and JavaScript scripting support extraction.
Findings and evidence
The Ubuntu 20.04.1/Linux 5.8.0/Docker 20.10.4 evaluation covers network-intensive, CPU-intensive, disk-I/O-intensive, and idle workloads. The CPU-intensive group has about 97.3% running/preempted time, the overall silhouette score is 0.6527, and the second clustering stage yields five finer groups. Average tracing overhead is 21.62% with all kernel events and 3.6% with the minimal necessary event set.
Limitations and future directions
Limitations: The evaluation covers four workload classes on one Linux/Docker configuration and depends on the selected execution-state features and clustering choices; it does not establish cross-runtime or multi-tenant generalization.
Future work: Use the clusters for resource allocation and configuration, test other inputs and tasks, distinguish interrupt types, and explore richer or deep-learning-based characterization.
Sources and identifiers
- Published version published
When to cite this paper
Cite this paper when characterizing container workloads from host-level tracing without internal container agents.
- LTTng pid_ns/tid context for attributing host events to containers.
- PageRank thread selection, execution-state vectors, and two-stage K-Means clustering.
- Docker workload groups, silhouette score, and minimal-versus-all-kernel-event overhead comparison.
Citation
@inproceedings{ezzatiJivan2021containerworkload,
author = {Madeline Janecek and Naser Ezzati-Jivan and Seyed Vahid Azhari},
title = {Container Workload Characterization Through Host System Tracing},
year = {2021},
booktitle = {IEEE International Conference on Cloud Computing Technology and Science (CloudCom)},
pages = {9-19},
publisher = {IEEE},
doi = {10.1109/ic2e52221.2021.00015},
url = {https://doi.org/10.1109/ic2e52221.2021.00015}
}Other citation formats for Word and reference managers
Janecek, M., Ezzati-Jivan, N., & Azhari, S. V. (2021). Container Workload Characterization Through Host System Tracing. In IEEE International Conference on Cloud Computing Technology and Science (CloudCom) (pp. 9-19). https://doi.org/10.1109/ic2e52221.2021.00015M. Janecek, N. Ezzati-Jivan, and S. V. Azhari, "Container Workload Characterization Through Host System Tracing," in IEEE International Conference on Cloud Computing Technology and Science (CloudCom), pp. 9-19, 2021, doi: 10.1109/ic2e52221.2021.00015