Energy-Efficient QoS-Aware Scheduling for S-NUCA Many-Cores

Sudam Maduranga Wasala, Jurre Wolff, Yixian Shen, Anuj Pathania, Clemens Grelck, Andy D. Pimentel

Abstract

Modern many-core processors increasingly utilize Static Non-Uniform Cache Access (S-NUCA) architectures, where the physical distance between a core and the shared last-level cache (LLC) results in varying access latencies. This spatial heterogeneity, combined with the inadequacy of traditional performance metrics like Instructions Per Second (IPS) for measuring true Quality of Service (QoS), presents significant challenges for energy-efficient task scheduling. This paper presents a novel scheduling policy that utilizes "Application Heartbeats" to monitor real-time performance and employs a dual-strategy of Dynamic Voltage and Frequency Scaling (DVFS) and thread migration. By prioritizing QoS stability and exploiting the physical layout of the chip, the proposed system achieves up to 18.7% energy savings over existing state-of-the-art methods while maintaining strict performance targets.

Introduction

In the pursuit of higher performance, many-core architectures have moved toward S-NUCA designs. In these systems, cores located near the center of the silicon die enjoy lower latency to the shared cache banks compared to those on the periphery. While this design improves scalability, it introduces a "spatial heterogeneity" where the performance of a thread depends heavily on its physical location.

Furthermore, traditional hardware-level metrics such as IPS often fail to reflect the actual progress of an application. For instance, a thread stuck in a spin-lock may exhibit high IPS without contributing to the application's goals, such as frame rendering or data processing. To address these issues, we propose a scheduling framework that uses application-level feedback—Heartbeats—to guide energy-saving decisions without compromising QoS.

Methodology

Application Heartbeats and the Magic Timestamp

The core of the proposed system is the "Application Heartbeat" framework. Instead of relying on generic hardware counters, applications emit signals (heartbeats) at key progress points (e.g., completing a frame or a transaction). The scheduler monitors the Heart Rate (HR) against a predefined target. To ensure accuracy within the HotSniper many-core simulator, a "Magic Timestamp Module" was implemented to synchronize the host and simulated clocks, preventing temporal drift from skewing performance data.

The 5-State Reactive Controller

The scheduling policy operates through a reactive 5-state controller that categorizes the system's performance relative to the target HR:

  • States A & E (Critical Deviation): When the HR is significantly above or below the target, the system employs "macro" adjustments, including large frequency jumps and immediate thread migrations.
  • States B & D (Minor Deviation): When the HR is near the target, the system uses "micro" adjustments to fine-tune performance and prevent oscillation.
  • State C (Target Met): Once the application reaches the target HR range, the controller shifts its priority entirely to energy minimization.

Control Knobs: DVFS and Migration

The controller utilizes two primary mechanisms to manage performance:

  1. DVFS: Provides immediate, high-impact changes to performance by adjusting clock speeds.
  2. Thread Migration: Leverages the S-NUCA layout by moving threads toward the center of the chip to boost performance (via lower cache latency) or toward the edges to reduce power consumption when performance exceeds requirements.

Experimental Setup

The policy was evaluated using the HotSniper many-core simulator, modeling a 64-core S-NUCA architecture in an 8x8 grid. The researchers utilized the PARSEC benchmark suite to represent diverse workloads:

  • Blackscholes: A computationally intensive financial modeling application.
  • Canneal: A memory-intensive workload focused on chip design routing.
  • Dedup: A data compression workload with moderate resource demands.

Results and Discussion

Energy Efficiency and Stability

The proposed heartbeat-aware policy demonstrated a significant improvement in energy efficiency, achieving an 18.7% reduction in energy consumption compared to existing state-of-the-art schedulers like Hierarchical Power Management (HPM) and PCMiG.

A key finding was the stability of the QoS. While competing methods often suffered from "jitter"—overshooting the target or oscillating wildly—the 5-state controller allowed applications to reach their target HR and remain there consistently. This stability is attributed to the hierarchical priority system: the scheduler only pursues energy savings once the QoS target is secured.

Spatial Awareness as an Asset

The study confirmed that frequency adjustments have a more immediate effect on performance than migration. However, by treating the S-NUCA chip as a "performance gradient," the scheduler successfully used thread migration as a tool for long-term stability and energy optimization. Moving threads to the periphery during State C allowed for significant power reduction without the immediate performance drop-off that aggressive frequency scaling might cause.

Conclusion and Future Work

This research demonstrates that application-level feedback, combined with a spatially-aware scheduling policy, can significantly enhance the energy efficiency of many-core processors. By prioritizing QoS and utilizing the inherent heterogeneity of S-NUCA architectures, the proposed system provides a robust framework for modern high-performance computing.

Future work will focus on integrating thermal awareness to prevent localized hotspots and core throttling. Additionally, the researchers aim to transition the policy from simulation to physical hardware and explore predictive modeling using machine learning to anticipate workload shifts before they impact the Heart Rate.

Summary updated

BibTeX
@inproceedings{wasala2025energy,
  author = {Sudam Maduranga Wasala and Jurre Wolff and Yixian Shen and Anuj Pathania and Clemens Grelck and Andy D. Pimentel},
  title = {{Energy-Efficient QoS-Aware Scheduling for S-NUCA Many-Cores}},
  booktitle = {ISQED},
  year = {2025},
  url = {https://openreview.net/forum?id=V2jFutjVrl},
}
Figure