Imitation Learning-Based Performance-Power Trade-Off Uncore Frequency Scaling Policy for Multicore System.

Area of Science:

• Computer Architecture
• Power Management
• Machine Learning

Background:

• Uncore components significantly contribute to multicore processor power consumption.
• Optimizing uncore frequency scaling (UFS) is crucial for overall system power efficiency.

Purpose of the Study:

• To develop a novel imitation learning-based UFS policy for enhanced power efficiency.
• To optimize processor Performance Per Watt (PPW) rather than just power saving.

Main Methods:

• Implemented an imitation learning policy using the DAgger algorithm for online learning.
• Focused on fine-tuning an expert model using online aggregation data to improve efficiency.
• Shifted optimization target to Performance Per Watt (PPW).

Main Results:

• The proposed UFS policy demonstrated superior performance compared to advanced UFS policies on SPEC CPU2017 benchmarks.
• Achieved up to a 10% improvement in performance relative to performance-first policies.
• Maintained stable processor operation near optimal power efficiency even with unseen processor loads.

Conclusions:

• The novel imitation learning-based UFS policy effectively enhances multicore processor power efficiency.
• The policy generalizes well to unseen loads and maintains optimal power efficiency.
• This approach offers a significant improvement in processor Performance Per Watt.