Can We Speculate Running Application With Server Power Consumption Trace?

Area of Science:

• Computer Science
• Machine Learning
• Data Center Management

Background:

• Server application detection is crucial for data center energy monitoring.
• Time series classification methods, including deep learning, are actively researched.
• Existing distance measurements for time series have limitations.

Purpose of the Study:

• To propose a novel distance measurement for time series classification.
• To develop a hybrid algorithm combining nearest neighbor and LSTM for application detection.
• To enhance data center energy consumption monitoring and analysis.

Main Methods:

• Proposed a novel Local Time Warping (LTW) distance measurement.
• Developed a hybrid 1-nearest neighbor (1NN)-LTW and Long Short-Term Memory (LSTM) algorithm.
• Combined prediction probability vectors from 1NN-LTW and LSTM for classification.

Main Results:

• Local Time Warping (LTW) improved classification accuracy from 84% to 90% compared to Dynamic Time Warping (DTW).
• A linear version of LTW achieved performance similar to state-of-the-art DTW methods with faster runtime.
• The hybrid algorithm achieved up to 93% accuracy in power series classification.

Conclusions:

• The proposed LTW is competitive with existing DTW variants, demonstrating the benefit of its noncommutative feature.
• The hybrid algorithm effectively classifies power consumption series for server application detection.
• This research offers insights into time series distance measurement and hybrid deep learning models.