Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Accurate Performance Predictors for Edge Computing Applications

Panagiotis Giannakopoulos, Bart van Knippenberg, Kishor Chandra Joshi, Nicola Calabretta, George Exarchakos

TL;DR

This work tackles performance predictability in edge computing by automatically building RTT predictors that balance accuracy with low inference overhead under dynamic co-location and node heterogeneity. It introduces perfCorrelate to select highly correlated monitoring metrics and tsfresh-based feature extraction, building separate predictors for each application-node pair with both non-sequential and sequential models. The results show up to roughly 90% RTT-prediction accuracy and inference times below 1% of RTT, with robust handling of dynamic co-location and the ability to predict remaining RTT during execution. The approach offers a practical path toward improving resource utilization and real-time scheduling in edge environments, with future work focusing on runtime deployment and scheduler integration.

Abstract

Accurate prediction of application performance is critical for enabling effective scheduling and resource management in resource-constrained dynamic edge environments. However, achieving predictable performance in such environments remains challenging due to the co-location of multiple applications and the node heterogeneity. To address this, we propose a methodology that automatically builds and assesses various performance predictors. This approach prioritizes both accuracy and inference time to identify the most efficient model. Our predictors achieve up to 90% accuracy while maintaining an inference time of less than 1% of the Round Trip Time. These predictors are trained on the historical state of the most correlated monitoring metrics to application performance and evaluated across multiple servers in dynamic co-location scenarios. As usecase we consider electron microscopy (EM) workflows, which have stringent real-time demands and diverse resource requirements. Our findings emphasize the need for a systematic methodology that selects server-specific predictors by jointly optimizing accuracy and inference latency in dynamic co-location scenarios. Integrating such predictors into edge environments can improve resource utilization and result in predictable performance.

Accurate Performance Predictors for Edge Computing Applications

TL;DR

This work tackles performance predictability in edge computing by automatically building RTT predictors that balance accuracy with low inference overhead under dynamic co-location and node heterogeneity. It introduces perfCorrelate to select highly correlated monitoring metrics and tsfresh-based feature extraction, building separate predictors for each application-node pair with both non-sequential and sequential models. The results show up to roughly 90% RTT-prediction accuracy and inference times below 1% of RTT, with robust handling of dynamic co-location and the ability to predict remaining RTT during execution. The approach offers a practical path toward improving resource utilization and real-time scheduling in edge environments, with future work focusing on runtime deployment and scheduler integration.

Abstract

Accurate prediction of application performance is critical for enabling effective scheduling and resource management in resource-constrained dynamic edge environments. However, achieving predictable performance in such environments remains challenging due to the co-location of multiple applications and the node heterogeneity. To address this, we propose a methodology that automatically builds and assesses various performance predictors. This approach prioritizes both accuracy and inference time to identify the most efficient model. Our predictors achieve up to 90% accuracy while maintaining an inference time of less than 1% of the Round Trip Time. These predictors are trained on the historical state of the most correlated monitoring metrics to application performance and evaluated across multiple servers in dynamic co-location scenarios. As usecase we consider electron microscopy (EM) workflows, which have stringent real-time demands and diverse resource requirements. Our findings emphasize the need for a systematic methodology that selects server-specific predictors by jointly optimizing accuracy and inference latency in dynamic co-location scenarios. Integrating such predictors into edge environments can improve resource utilization and result in predictable performance.
Paper Structure (25 sections, 2 equations, 7 figures, 2 tables)

This paper contains 25 sections, 2 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. System monitoring
  5. Feature selection and correlation analysis
  6. Metric-Features Extraction
  7. Redundancy Removal
  8. Extract Correlations
  9. Predictor generation and evaluation
  10. Non-sequential Models
  11. Sequential Models
  12. Model selection
  13. Experimental set up
  14. Applications
  15. Frameworks
  16. ...and 10 more sections

Figures (7)

  • Figure 1: Overview of the proposed architecture for accurate performance prediction in edge computing environments. The system operates in three main stages: (1) System monitoring: The monitoring system continuously collects time-series data from each server, including metrics such as CPU, memory, network usage, and RTT values per task. This data is stored in a centralized database. (2) Feature selection and correlation analysis (perfCorrelate): From the complete set of monitoring metrics, time-series features are extracted. perfCorrelate computes the correlation between these features and historical RTTs to identify the most relevant metrics. Redundant features are removed to produce a compact and informative feature set. (3) Predictor generation and evaluation: Using the selected metrics, both sequential and non-sequential machine learning models are trained to predict task RTTs before execution. Sequential models use the raw time-series form of the metrics, while non-sequential models use the extracted features. All models are evaluated based on RMSE and inference time, and the best-performing one is selected for each application-node pair.
  • Figure 2: Structure of the used infrastructure. The servers are connected through 1 Gbps Ethernet link.
  • Figure 3: RMSE of non-sequential models for each application on Worker-3, showing how increasing the number of input features and history window length affects prediction accuracy. The plot also illustrates the RMSE when the mean RTT is used as the prediction.
  • Figure 4: RMSE of sequential models for each application on Worker-3, showing how increasing the number of input features and history window length affects prediction accuracy. The plot also illustrates the RMSE when the mean RTT is used as the prediction.
  • Figure 5: Training and inference times of sequential and non-sequential models for MotionCor2 on Worker-3.
  • ...and 2 more figures