Dynamic Energy and Thermal Management of Multi-core Mobile Platforms: A Survey

Amit Kumar Singh, Somdip Dey, Klaus D. McDonald-Maier, Basireddy Karunakar Reddy, Geoff V. Merrett, Bashir M. Al-HashimiIEEE Des. Test 2020 · DOI

Abstract

Modern mobile devices face a critical challenge: balancing the demand for high-performance computing with strict battery life constraints and thermal safety limits. This paper surveys state-of-the-art Dynamic Energy and Thermal Management (DETM) techniques, categorizing them into proactive and reactive strategies. It specifically addresses the complexities introduced by heterogeneous multi-core processors, such as ARM’s big.LITTLE architecture, and explores how intelligent resource management can mitigate the "thermal wall" in smartphones and wearables.

Introduction

As mobile applications evolve to include high-end gaming, augmented reality (AR/VR), and on-device AI, hardware has shifted from single-core processors to complex heterogeneous Multi-Processor Systems-on-Chips (MPSoCs). While these architectures offer significant theoretical performance, they are constrained by a "thermal wall." Simply increasing clock speeds leads to excessive heat and rapid battery depletion, while aggressive thermal throttling can severely degrade the user experience.

The core challenge lies in Dynamic Resource Management (DRM): the ability of an operating system to intelligently assign tasks to appropriate cores and adjust voltages and frequencies in real-time to maximize efficiency without exceeding thermal limits.

System Architecture for Resource Management

Effective DETM requires a multi-layered approach that bridges the gap between high-level application demands and low-level hardware constraints. A typical framework involves a Run-Time Manager (RTM) situated within the Operating System (OS) layer.

A high-level overview of a Dynamic Energy and Thermal Management (DETM) framework, illustrating the interaction between applications, the run-time manager (OS/RTM), and hardware resources.

As shown in the figure above, the RTM utilizes various inputs—including temperature estimation models, workload estimation models, and real-time temperature sensors—to make informed decisions. These decisions result in hardware resource selection and Voltage/Frequency (V/F) control. In return, the hardware provides energy and performance statistics (e.g., hardware counters) to the RTM to close the feedback loop.

Management Taxonomies

The survey classifies management approaches based on two primary dimensions:

1. Objective

  • Dynamic Energy Management (DEM): Primarily focuses on extending battery life through techniques like Dynamic Voltage and Frequency Scaling (DVFS) and Dynamic Power Management (DPM).
  • Dynamic Thermal Management (DTM): Focuses on avoiding localized "hotspots" and spatial temperature gradients that can cause physical degradation of the silicon over time.

2. Control Mechanism

  • Reactive: These methods trigger only when a specific threshold is reached (e.g., throttling a CPU when it hits 80°C). While low-overhead, they are often "too late" to prevent performance dips.
  • Proactive: These strategies use predictive models—such as Machine Learning, Linear Regression, or Integer Linear Programming—to estimate future workloads or temperatures. By taking preemptive action, they optimize the system before a thermal violation occurs.

Key Insights and Results

  • The Reliability Link: Thermal management is vital for hardware longevity. A 10°C to 20°C increase in operating temperature can decrease the lifetime reliability of a chip's metallic structures by up to 16 times.
  • AI-Driven Classification: A growing trend involves using Convolutional Neural Networks (CNNs) to classify program source code at runtime (e.g., as compute-intensive or memory-intensive) to guide power management decisions.
  • Hybrid Superiority: The survey finds that hybrid task-mapping algorithms—combining offline profiling with online adaptation—generally outperform pure strategies by leveraging design-time accuracy with runtime flexibility.
  • Imitation Learning: Frameworks using imitation learning for core and frequency control have demonstrated the ability to generate near-optimal policies with significantly less overhead than traditional Reinforcement Learning (RL).

Future Directions

The survey identifies several emerging frontiers in mobile resource management:

  • Hierarchical Management: As chips move toward "multi-cluster" designs with three or more core types, centralized managers may become bottlenecks, necessitating local "sub-managers" for each cluster.
  • Multi-objective Optimization: Future managers must balance energy and heat alongside security (to prevent side-channel attacks) and long-term reliability.
  • Cloud Offloading: With 5G, the decision to process a task locally or offload it to the cloud is becoming a critical component of the energy-management equation.
  • User-Centric Metrics: There is a shift toward optimizing for "User Experience" (UX) rather than just raw throughput, requiring new mathematical models to quantify subjective performance.

Summary updated

BibTeX
@article{singh2020dynamic,
  author = {Amit Kumar Singh and Somdip Dey and Klaus D. McDonald-Maier and Basireddy Karunakar Reddy and Geoff V. Merrett and Bashir M. Al-Hashimi},
  title = {{Dynamic Energy and Thermal Management of Multi-core Mobile Platforms: A Survey}},
  journal = {IEEE Des. Test},
  year = {2020},
  doi = {10.1109/mdat.2020.2982629},
}
Figure