Intelligent Performance Analysis

Objectives

• Develop performance analysis techniques for online streaming data.
• Develop online anomaly detection and pattern searching techniques.
• Develop deep learning techniques for performance prediction.

Progress and Accomplishments

• Developed prediction methods for network data throughput and cluster job performance
• Machine leaning based model
• Statistical model - Best Predictive Generalized Linear Mixed Model with Predictive Lasso (IJSP5/2015 journal paper)
• Time series model (ICNC'15 conf. paper)
• Developed automated performance diagnostic method and failure detection method for cluster jobs
• Developed a data reduction method based on pattern searching - U.S. Patent pending serial no. 14/555,365.
• Developing in-situ feature detection methods
• Developing online job failure prediction methods

Impacts

• Improve resource utilization and analysis workflow performance.
• Enable predictable data throughput over the network.
• Improve resilience of computing jobs
• Long-term capacity and traffic engineering planning of network infrastructures
• Enable dynamic data reduction based on in-situ feature detection