Learning Cache Coherence Traffic for NoC Routing Design

Guochu Xiong, Xiangzhong Luo, Weichen Liu

TL;DR

This paper introduces a Deep Reinforcement Learning (DRL) framework that jointly optimizes Network-on-Chip (NoC) topology and routing by explicitly accounting for cache coherence traffic. By using a novel analyzer (CCTA) to bridge the gap between memory consistency protocols and network design, the authors achieve up to 55% faster execution times and 49% energy savings compared to traditional methods.

What Problem Does This Paper Address?

In modern multi-core processors, the Network-on-Chip (NoC) is the highway system for data. To keep data consistent across multiple caches, systems use cache coherence protocols (like MESI). These protocols generate a massive amount of "background" traffic—requests to update states, invalidation signals, and data fetches—that traditional NoC routing designs often ignore.

Most existing NoC routing algorithms are designed and tested using "synthetic traffic" (random or patterned data bursts) that fails to capture the nuances of real-world cache-to-cache communication. This leads to a "design-evaluation mismatch": a routing strategy might look great in a simulator but perform poorly in a real system because it didn't account for the congestion caused by coherence messages. Furthermore, routing and network topology (how the cores are physically connected) are often treated as separate problems, leading to suboptimal performance.

Methods

The authors propose a holistic approach called Case 3 design, which integrates cache coherence into both the design and evaluation phases. Their framework consists of two main pillars:

  1. Cache Coherence Traffic Analyzer (CCTA): A new tool integrated with the Gem5 simulator. It tracks the lifecycle of coherence messages, marking start and end points within the L1 cache to measure "CPU delay" and "write miss time." This provides the granular data needed to understand how coherence impacts the network.
  2. DRL-based Joint Optimization: The system uses Deep Reinforcement Learning to make two decisions simultaneously:
  3. Topology Selection: A Q-network evaluates candidate structures (like Mesh, Torus, or Fat-Tree) based on current network and coherence states.
  4. Dynamic Routing: A three-layer Multi-Layer Perceptron (MLP) updates "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.

Datasets

The researchers evaluated their framework using the PARSEC 2.1 benchmark suite, which contains diverse, real-world multi-core workloads (e.g., X264 for video encoding, Vips for image processing, and Fluidanimate for physical simulations).

Simulations were conducted on:

  • 16-core and 64-core configurations.
  • Topologies: Crossbar, Mesh, Point-to-Point (Pt2Pt), Torus, Fat-Tree, and Flattened Butterfly.
  • Simulator: Gem5 with the Garnet NoC model and McPAT for energy estimation.

Key Insights and Ideas

  • Coherence is a Dual-Role Player: Routing isn't just about moving data from A to B; it's about facilitating data sharing during computation. Efficient cache-to-cache sharing (enabled by coherence) actually reduces the time a CPU spends waiting for data from main memory.
  • The "Aha" Moment on Topology: The authors found that routing performance is inherently tied to network structure. By allowing the DRL agent to pick the topology and the routing paths, the system can adapt to high-congestion scenarios by switching to topologies with more routing paths, even if they aren't "standard" for that core count.
  • Closing the Gap: Traditional "non-cc-aware" (non-cache-coherence-aware) models consistently underestimate energy usage and latency because they miss the "bursty" nature of coherence traffic.

Key Results

The proposed method significantly outperformed standard routing (XY) and other RL-based approaches:

  • Latency: Achieved up to 10.52% lower packet latency compared to the congestion-adaptive DyAD algorithm.
  • Execution Speed: Workloads finished up to 55.51% faster than with standard cache-coherence-aware RL.
  • Energy Efficiency: Total energy savings reached 49.02% compared to XY routing.
  • Scalability: The performance gains were even more pronounced in the 64-core system, where the complexity of coherence traffic is much higher, proving the framework's robustness for large-scale systems.

Limitations and Future Work

While the paper demonstrates impressive results, a few areas remain for exploration:

  • Protocol Variety: The study focused primarily on the MESI protocol. Future work could examine how more complex protocols (like MOESI or directory-based schemes for heterogeneous chips) affect the DRL agent's learning.
  • Training Overhead: DRL can be computationally expensive to train. While the authors provide pretrained models, the real-time overhead of the "WeightPredictor" MLP in a hardware implementation isn't fully detailed.
  • Heterogeneity: The current model assumes a relatively uniform multi-core environment; adapting this to heterogeneous SoCs (mixing CPUs, GPUs, and NPUs) would be a logical next step.

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},
}

Loading related content...

Figure