Efficient Detection of Communication-related Performance Anti-patterns in Microservices

Catalog abstract summary

The preprint combines selected communication-related system-call traces with distributed observability traces to detect Blob and Empty Semi Trucks anti-patterns in microservices. It reports supervised, semi-supervised, and unsupervised settings on DeathStarBench scenarios.

Source: Authorea public preprint abstract, paraphrased; the preprint is not peer-reviewed.

Problem and motivation

Communication-related performance antipatterns in microservices can be hidden across user-level request traces and lower-level communication activity. The paper targets detection of Blob and Empty-semi-trucks while limiting collection overhead.

Method and contribution

The workflow combines distributed observability traces with selected communication-related system calls, correlates user-space and kernel events, aggregates features, and supports supervised, semi-supervised, and unsupervised learning. The reported setup uses LTTng, Trace Compass, Jaeger-client instrumentation, liblttng-ust, and kernel tracing, with offline training and online detection.

Findings and evidence

The evaluation uses DeathStarBench with 14 clean/noisy scenarios. The abstract reports up to 91% accuracy, 2.74% data-collection overhead, and more than 80% agreement with manual analysis. The later discussion reports 63% for the unsupervised setting and higher supervised performance, with semi-supervised learning positioned as a labeling-cost/accuracy compromise.

Limitations and future directions

Limitations: Evidence is a remote-read public preprint, not a locally hashed or peer-reviewed version. The approach depends on the LTTng/Jaeger/Trace Compass stack and communication-specific calls such as recvfrom/recvmsg/recvmmsg and sendto/sendmsg/sendmmsg. It is evaluated on controlled DeathStarBench scenarios, not proprietary industrial systems. Tool versions, OS/kernel, hardware, runtime, and replication details are unknown in the retained evidence note.

Future work: Evaluate with industrial partners, add I/O-related system-call families, and investigate topic-modeling or generative-AI extensions.

Sources and identifiers

• Published version preprint
• Public preprint PDF · PDF public_full_text

When to cite this paper

Cite this paper when your work uses or compares combining distributed spans with communication-related kernel/user system calls to detect microservice performance antipatterns.

• combining distributed spans with communication-related kernel/user system calls to detect microservice performance antipatterns.
• the Blob and Empty-semi-truck target classes and the 14-scenario DeathStarBench evaluation.
• the supervised/semi-supervised/unsupervised comparison and its reported collection-overhead/manual-agreement measurements.
• with an explicit preprint and controlled-testbed caveat; do not present it as peer-reviewed industrial evidence.

Citation

@misc{ezzatiJivan2025efficientdetection,
  author = {Masoumeh Nourollahi and Naser Ezzati-Jivan and Adel Belkheiri and Michel Dagenais},
  title = {Efficient Detection of Communication-related Performance Anti-patterns in Microservices},
  year = {2025},
  howpublished = {Authorea preprint},
  publisher = {Wiley},
  doi = {10.22541/au.175533132.24109345/v1},
  url = {https://doi.org/10.22541/au.175533132.24109345/v1}
}

Nourollahi, M., Ezzati-Jivan, N., Belkheiri, A., & Dagenais, M. (2025). Efficient Detection of Communication-related Performance Anti-patterns in Microservices. Authorea preprint. https://doi.org/10.22541/au.175533132.24109345/v1

M. Nourollahi, N. Ezzati-Jivan, A. Belkheiri, and M. Dagenais, "Efficient Detection of Communication-related Performance Anti-patterns in Microservices," in Authorea preprint, 2025, doi: 10.22541/au.175533132.24109345/v1