BIOS-Based Server Intelligent Optimization.

Area of Science:

• Computer Science
• Systems Engineering

Background:

• Server performance tuning is crucial for enterprise applications but often lacks systematic approaches.
• Application-layer tuning requires specific knowledge and is not always efficient.
• Existing dynamic tuning methods use predictive models for hardware prefetching.

Purpose of the Study:

• To develop a BIOS-based dynamic tuning framework for Taishan 2280 servers.
• To implement a joint BIOS optimization algorithm using a deep Q-network for efficient configuration management.
• To enhance server performance and stability through intelligent, adaptive tuning.

Main Methods:

• Designed a BIOS dynamic tuning framework with dynamic identification and static optimization.
• Utilized performance monitor counters (PMCs) for workload scenario recognition.
• Developed a deep Q-network algorithm, modeled as a Markov decision process, incorporating state machine control for optimization.

Main Results:

• The proposed algorithm improved performance up to 1.10× compared to experience-based configurations.
• Demonstrated superior performance over genetic algorithms and particle swarm optimization.
• Reduced the probability of server downtime and generated fewer infeasible solutions than heuristic algorithms.

Conclusions:

• The dynamic tuning framework is extensible and adaptable to various scenarios and server configurations.
• The deep Q-network-based BIOS optimization offers a systematic and efficient approach to server performance tuning.
• This method provides significant performance gains and enhanced stability for enterprise servers.