MemAdapt: Adaptive Monitoring of Memory Usage Through Irregularly Sampled Data

Catalog abstract summary

The available author synopsis describes an adaptive memory-monitoring approach built around irregularly sampled data and forecasting; the full algorithm and evaluation were not captured.

Source: OpenAlex abstract metadata and author synopsis reviewed; the publisher abstract is not reproduced because reuse permission was not established.

Problem and motivation

Memory monitors often produce irregular observations, whereas many forecasting methods assume regular time steps. Fixed-rate sampling can oversample stable periods, undersample rapid changes, and waste monitoring resources.

Method and contribution

The study derives features from sar, including %memusg, %swpused, %hugused, %user, and the inverse of %vmeff, and uses approximately 85,000 timestamps from one day. It evaluates irregular subsampling with a 900-second horizon, lookback values n=3 through 7, adaptive rates from 0.5 to 30 seconds, thresholds m=0.8 and 0.5, and theta=0.5. ODE-RNN, RNN, Kalman Filter, KF-LSTM, and related models are evaluated with an 85/15 split under sysbench/stress workloads and MSE/error metrics.

Findings and evidence

ODE-RNN is the strongest model in the reported comparisons. The paper reports 78.5% of sampling-rate predictions within 5% of the ideal rate and approximately 82.5% and 83.4% within 10% and 15%, respectively. At n=7, the reported MSE comparison is 1.768 for ODE-RNN versus 1.997 for a comparison configuration. Longer history improves prediction in the evaluated traces, supporting adaptive rather than fixed-rate monitoring.

Limitations and future directions

Limitations: The data are derived from one day and selected workloads, so they do not establish behavior under long-running production drift, abrupt pressure events, multiple applications, or heterogeneous machines. The operational overhead and control-loop stability of deploying the adaptive policy are not fully established by the offline evaluation.

Future work: Evaluate months-long and multi-host traces, abrupt memory-pressure and swap storms, online retraining, sampling-policy stability, and end-to-end CPU/storage/network overhead in production-like monitoring agents.

Sources and identifiers

• Published version published
• CASCON program public_source_record

When to cite this paper

Cite this paper when adapting memory-monitoring frequency from irregular observations and usage forecasts.

• ODE-RNN and alternative irregular-time forecasting models.
• The forecast-to-sampling-rate control policy.
• The reported within-5/10/15-percent sampling-rate accuracy.
• The one-day offline-data and production-deployment boundary.

Citation

@inproceedings{ezzatiJivan2024memadaptadaptive,
  author = {Pranjal Chakraborty and Majid Babaei and Leila Tahmooresnejad and Naser Ezzati-Jivan},
  title = {MemAdapt: Adaptive Monitoring of Memory Usage Through Irregularly Sampled Data},
  year = {2024},
  booktitle = {2024 IEEE International Conference on Collaborative Advances in Software and COmputiNg (CASCON)},
  pages = {1-6},
  publisher = {IEEE},
  doi = {10.1109/CASCON62161.2024.10838037},
  url = {https://doi.org/10.1109/CASCON62161.2024.10838037}
}

Chakraborty, P., Babaei, M., Tahmooresnejad, L., & Ezzati-Jivan, N. (2024). MemAdapt: Adaptive Monitoring of Memory Usage Through Irregularly Sampled Data. In 2024 IEEE International Conference on Collaborative Advances in Software and COmputiNg (CASCON) (pp. 1-6). https://doi.org/10.1109/CASCON62161.2024.10838037

P. Chakraborty, M. Babaei, L. Tahmooresnejad, and N. Ezzati-Jivan, "MemAdapt: Adaptive Monitoring of Memory Usage Through Irregularly Sampled Data," in 2024 IEEE International Conference on Collaborative Advances in Software and COmputiNg (CASCON), pp. 1-6, 2024, doi: 10.1109/CASCON62161.2024.10838037