Learning Cache Coherence Traffic for NoC Routing Design

Guochu Xiong, Xiangzhong Luo, Weichen Liu

Abstract

The rapid growth of multi-core systems highlights the need for efficient Network-on-Chip (NoC) design to ensure seamless communication. Cache coherence, essential for data consistency, substantially reduces task computation time by enabling data sharing among caches. As a result, routing serves two roles: facilitating data sharing and managing NoC-level communication. However, cache coherence is often overlooked in routing, causing mismatches between design expectations and evaluation outcomes. In this work, we propose a cache coherence-aware routing approach with integrated topology selection, guided by our Cache Coherence Traffic Analyzer (CCTA). Our method achieves up to 10.52% lower packet latency, 55.51% faster execution time, and 49.02% total energy savings, underscoring the critical role of cache coherence in NoC design and enabling effective co-design.

Introduction

In modern multi-core processors, the Network-on-Chip (NoC) acts as the primary communication fabric. As core counts scale, the efficiency of this fabric becomes the primary determinant of system performance. Central to this communication is the cache coherence protocol (e.g., MESI), which ensures data consistency across distributed caches. While these protocols are vital for shared-memory programming, they generate significant "background" traffic—including invalidation signals and state-update requests—that traditional NoC routing designs often ignore.

Historically, NoC routing algorithms have been designed and validated using synthetic traffic patterns. However, these patterns fail to capture the bursty, interdependent nature of real-world cache-to-cache communication. This leads to a "design-evaluation mismatch," where a routing strategy optimized for random traffic underperforms in a real system because it cannot handle the specific congestion patterns caused by coherence messages.

The Challenge: Coherence-Aware Design

Two primary challenges hinder the development of optimized NoCs for multi-core systems. First, there is a lack of specialized tools to isolate and assess how cache coherence impacts network performance. Second, routing and network topology (the physical layout of connections) are often treated as separate problems, leading to suboptimal performance when the network is under heavy load.

Efficient routing must serve a dual role: it must facilitate the data sharing required by computation while simultaneously managing the physical communication constraints of the NoC. When cache-to-cache sharing is handled efficiently, it reduces the time a CPU spends waiting for data from main memory, directly impacting execution speed.

Methodology: The CCTA and DRL Framework

To address these challenges, we propose a holistic framework that integrates cache coherence into both the design and evaluation phases. The framework consists of two main pillars: the Cache Coherence Traffic Analyzer (CCTA) and a Deep Reinforcement Learning (DRL) optimization engine.

Cache Coherence Traffic Analyzer (CCTA)

The CCTA is a novel tool integrated with the Gem5 simulator. Unlike standard network monitors, it tracks the entire lifecycle of coherence messages by marking start and end points within the L1 cache. This allows the tool to measure "CPU delay" and "write miss time" directly attributable to network conditions, providing the high-fidelity data necessary to train intelligent routing agents.

DRL-based Joint Optimization

Our framework utilizes a DRL agent to perform joint optimization of topology and routing simultaneously:

  • Topology Selection: A Q-network evaluates candidate structures—including Mesh, Torus, Fat-Tree, and Flattened Butterfly—based on the current network state and coherence traffic intensity.
  • Dynamic Routing: A three-layer Multi-Layer Perceptron (MLP) acts as a WeightPredictor, updating link weights to steer traffic away from congestion in real-time.

The reward function for the DRL agent is multi-objective, balancing packet latency, CPU delay, and the total volume of coherence messages. This ensures the system optimizes for overall execution speed rather than just raw throughput.

Experimental Results

We evaluated the framework using the PARSEC 2.1 benchmark suite, which represents diverse, real-world multi-core workloads. Simulations were conducted on 16-core and 64-core configurations using the Garnet NoC model and McPAT for energy estimation.

The proposed coherence-aware approach significantly outperformed standard routing (XY) and existing congestion-adaptive algorithms (DyAD):

  • Execution Speed: Workloads completed up to 55.51% faster than with standard cache-coherence-aware RL models, proving that joint topology-routing optimization is superior to optimizing routing alone.
  • Energy Efficiency: The framework achieved 49.02% total energy savings compared to XY routing, primarily by reducing the time cores spend idling while waiting for coherence transitions.
  • Latency Reduction: We observed up to 10.52% lower packet latency, particularly in high-traffic scenarios where the DRL agent could adapt the topology to provide more efficient routing paths.

Conclusion

This research demonstrates that accounting for cache coherence is critical for modern NoC design. By bridging the gap between memory consistency protocols and network architecture via the CCTA and DRL-based co-design, we can achieve substantial gains in performance and energy efficiency. Future work will explore the impact of more complex coherence protocols and the adaptation of this framework for heterogeneous SoCs mixing CPUs, GPUs, and NPUs.

Summary updated

Abstract

The rapid growth of multi-core systems highlights the need for efficient Network-on-Chip (NoC) design to ensure seamless communication. Cache coherence, essential for data consistency, substantially reduces task computation time by enabling data sharing among caches. As a result, routing serves two roles: facilitating data sharing (influenced by topology) and managing NoC-level communication. However, cache coherence is often overlooked in routing, causing mismatches between design expectations and evaluation outcomes. Two main challenges are the lack of specialized tools to assess cache coherence's impact and the neglect of topology selection in routing. In this work, we propose a cache coherence-aware routing approach with integrated topology selection, guided by our Cache Coherence Traffic Analyzer (CCTA). Our method achieves up to 10.52% lower packet latency, 55.51% faster execution time, and 49.02% total energy savings, underscoring the critical role of cache coherence in NoC design and enabling effective co-design.
BibTeX
@inproceedings{xiong2025learning,
  author = {Guochu Xiong and Xiangzhong Luo and Weichen Liu},
  title = {{Learning Cache Coherence Traffic for NoC Routing Design}},
  booktitle = {ACM Great Lakes Symposium on VLSI},
  year = {2025},
  url = {https://openreview.net/forum?id=bICABoMWlo},
}
Figure