Modeling and Controlling Many-Core HPC Processors: an Alternative to PID and Moving Average Algorithms

Giovanni Bambini, Alessandro Ottaviano, Christian Conficoni, Andrea Tilli, Luca Benini, Andrea Bartolini

Abstract

Modern many-core High-Performance Computing (HPC) processors face significant thermal and power management challenges due to non-linear leakage effects and actuator coupling across shared voltage domains. Traditional control strategies, such as Proportional-Integral-Derivative (PID) and Moving Average algorithms, often struggle to maintain thermal constraints without sacrificing performance. This paper presents a Fuzzy-inspired Iterative Control Algorithm (FCA) that utilizes a heuristic look-up table for thermal capping and a bisection-based iterative solver for voltage selection. Validated via Hardware-in-the-Loop (HIL) simulation on a RISC-V microcontroller, the FCA reduces thermal limit violations by up to 5.52x while improving application execution speed by an average of 3.56% compared to industry-standard voting box algorithms.

Introduction

As semiconductor manufacturing pushes the limits of integration, the density of cores on a single silicon die has transformed thermal management into a complex Multiple-Input Multiple-Output (MIMO) control problem. In the context of HPC, maintaining a strict thermal envelope is critical to prevent hardware degradation and thermal runaway. However, existing industry solutions—such as Intel’s RAPL or AMD’s SMU—frequently rely on linear control theory. These methods often fail to account for the exponential relationship between temperature and static power leakage or the constraints imposed by shared voltage rails. This paper proposes an alternative control framework designed to handle these non-linearities with greater precision and efficiency.

The Challenges of Modern Many-Core Control

Two primary hurdles complicate the control of high-density processors:

  1. Exponential Leakage Feedback: In sub-10nm nodes, higher temperatures lead to an exponential increase in static power leakage. This creates a feedback loop where heat generates more power consumption, which in turn generates more heat. Linear controllers like PID are often too slow to react to these shifts or prone to oscillations.
  2. Actuator Coupling: To minimize hardware overhead, multiple cores typically share a single voltage regulator module (VRM). This "voltage domain" coupling means that if a single core requires a high voltage for a demanding task, all other cores in that domain are forced to operate at that higher voltage, leading to wasted power and unnecessary thermal overhead.

Methodology: The Fuzzy-inspired Iterative Control Algorithm (FCA)

The proposed FCA framework is built upon three technical pillars designed to address the shortcomings of linear controllers.

Advanced Thermal and Power Modeling

The authors developed a high-fidelity model that incorporates real-world hardware non-idealities. This includes the transition delays associated with Phase-Locked Loops (PLLs) and VRMs, alongside a rigorous accounting of the temperature-dependent nature of leakage power. This model serves as the foundation for the controller's decision-making process.

Fuzzy Thermal Capping

Rather than relying on a standard PID loop, the FCA employs a heuristic Look-Up Table (LUT) inspired by fuzzy control theory. By making "if-then" decisions based on both the current temperature and its derivative (the rate of change), the controller can anticipate thermal excursions. This model-free approach allows the system to manage the non-linear thermal-power feedback loop more gracefully than traditional linear methods.

Iterative Root-Finding for Voltage Selection

To address the problem of shared voltage domains, the authors replaced the standard "Voting Box Algorithm" (which simply selects the maximum requested voltage) with a bisection-based iterative solver. This method treats voltage selection as a root-finding problem, identifying an optimal operating point that balances the performance requirements of all cores in a domain while strictly adhering to the total power budget.

Experimental Setup and Evaluation

The system was validated using a Hardware-in-the-Loop (HIL) simulation environment. The control algorithm was implemented on a 50MHz RISC-V embedded microcontroller (the Low-Level Controller) mapped onto a Xilinx Zynq FPGA. The evaluation utilized three distinct synthetic workload scenarios:

  • MAX-WL: Heavy vector workloads across all cores.
  • MULTI-WL: A heterogeneous mix of heavy, light, and idle cores to test voltage coupling.
  • CLOUD-WL: Randomized workloads simulating unpredictable cloud environments.

These workloads were tested across three cooling configurations: WATER (homogeneous), AIR (Gaussian distribution), and RACK (column-wise distribution).

Results

The FCA was compared against an Enhanced Baseline (EBA) and the industry-standard Voting Box Algorithm (VBA). The results demonstrate significant improvements in both safety and performance:

  • Thermal Safety: The FCA achieved a maximum exceeded temperature 5.52x lower than the VBA and 2.44x lower than the EBA, effectively eliminating the risk of thermal runaway.
  • Performance Gains: Despite the stricter thermal management, the FCA improved application execution speed by an average of 3.56% across all tested scenarios.
  • Robustness: The controller showed the lowest performance variation across different cooling models (Air vs. Water), suggesting it is a highly robust solution for diverse data center environments.
  • Computational Efficiency: The algorithm is lightweight, occupying only ~78% of the 500µs control hyperperiod on a modest 50MHz RISC-V core.

Conclusion

The Fuzzy-inspired Iterative Control Algorithm represents a significant departure from traditional linear control in HPC power management. By acknowledging the non-linearities of leakage and the complexities of core coupling, the FCA provides a safer and more performant alternative to PID-based systems. Future work will focus on refining workload estimation and exploring the feasibility of Model Predictive Control (MPC) within the resource constraints of embedded microcontrollers.

Summary updated

Abstract

The race towards performance increase and computing power has led to chips with heterogeneous and complex designs, integrating an ever-growing number of cores on the same monolithic chip or chiplet silicon die. Higher integration density, compounded with the slowdown of technology-driven power reduction, implies that power and thermal management become increasingly relevant. Unfortunately, existing research lacks a detailed analysis and modeling of thermal, power, and electrical coupling effects and how they have to be jointly considered to perform dynamic control of complex and heterogeneous mpsoc. To close the gap, in this work, we first provide a detailed thermal and power model targeting a modern hpc mpsoc. We consider real-world coupling effects such as actuators’ non-idealities and the exponential relation between the dissipated power, the temperature state, and the voltage level in a single processing element. We analyze how these factors affect the control algorithm behavior and the type of challenges that they pose. Based on the analysis, we propose a thermal capping strategy inspired by Fuzzy control theory to replace the state-of-the-art PID controller, as well as a root-finding iterative method to optimally choose the shared voltage value among cores grouped in the same voltage domain. We evaluate the proposed controller with model-in-the-loop and hardware-in-the-loop co-simulations. We show an improvement over state-of-the-art methods of up to \(5\times\) the maximum exceeded temperature while providing an average of \(3.56\%\) faster application execution runtime across all the evaluation scenarios.
BibTeX
@article{bambini2024modeling,
  author = {Giovanni Bambini and Alessandro Ottaviano and Christian Conficoni and Andrea Tilli and Luca Benini and Andrea Bartolini},
  title = {{Modeling and Controlling Many-Core HPC Processors: an Alternative to PID and Moving Average Algorithms}},
  journal = {ACM Transactions on Autonomous and Adaptive Systems},
  year = {2024},
  doi = {10.1145/3694687},
}
Figure