Disaggregated Memory at the Edge

Luis M. Vaquero, Yehia Elkhatib, Félix CuadradoarXiv 2021 · arXiv

Abstract

This paper explores the application of "memory disaggregation"—a technique typically reserved for high-end data centers—to the network edge. By leveraging Non-Volatile Memory (NVM) and low-power wireless mesh networks, the authors create a dynamic, shared virtual memory space. This framework allows mobile devices, such as autonomous vehicles, to offload state and reduce local computational burdens by up to 60%.

Introduction

In traditional computing, memory is tightly coupled to a specific CPU. Resource disaggregation breaks this bond, allowing a pool of memory to be shared across many servers. While common in data centers using high-speed wired interconnects, the Edge presents unique challenges: devices are mobile, battery-powered, and connect over volatile wireless links.

Current edge devices (IoT sensors, drones, autonomous vehicles) struggle with limited local resources. If they require more data than they can hold, they typically rely on distant cloud services, which introduces significant latency. This research tackles the challenge of creating a "seamlessly integrated edge memory" that follows a device as it moves, providing low-latency access to data stored in the surrounding physical environment.

Proposed Framework: Edge-Level Disaggregation

The authors introduce a framework designed to decouple memory from local processing at the edge. The architecture consists of several key layers that bridge the gap between physical memory modules and mobile clients.

Figure 2: Proposed architecture for edge-level memory disaggregation, showing the interaction between the CPU, memory modules, and the aerial communication layer.

Key components include:

  • Flattened Communication Stack: Instead of heavy network protocols, the system encapsulates classic memory "load/store" primitives directly into wireless protocols like Zigbee (802.15.4).
  • NVM Mesh/Overlay: Non-Volatile Memory (NVM) is used for its persistence—crucial for volatile edge environments. These modules form a self-organizing mesh.
  • Hardware Memory Interface: A custom interface based on Gen-Z and Redfish standards allows memory modules to act as routers, forwarding requests to the correct physical location in the mesh.
  • Dynamic Federation: As a mobile device moves, it "joins" the local memory overlay of nearby objects (e.g., smart lamp posts), expanding its virtual memory address space on the fly.

Experimental Evaluation

The approach was evaluated using two custom-built physical prototypes:

  1. Smart Building Setup: A Raspberry Pi-based system controlling color-configurable bulbs to test remote configuration persistence.
  2. Autonomous Vehicle Simulation: A 1/10th scale self-driving car track featuring 30 Raspberry Pis acting as "lamp posts" equipped with SD cards (simulating NVM) to store environmental data.
Figure 6: The autonomous vehicle simulation track, where NVM-equipped nodes (lamp posts) provide a shared memory space for the moving vehicle.

Key Results and Insights

  • Computational Efficiency: In the car simulation, the reinforcement learning algorithm processed 60% less information by reading static road properties directly from edge memory rather than sensing them in real-time.
  • Performance: The custom protocol delivered higher throughput and greater stability for memory-centric tasks compared to traditional TCP over WiFi.
  • Latency: For small payloads (under 200 bytes), latency remained under 200ms even across five "hops" in the memory mesh.
  • Mobility Limits: The prototype functioned well at low speeds but struggled to maintain connections above 6 km/h, as the protocol could not converge fast enough to include the moving car in the mesh.

Limitations and Future Work

The current system's primary constraint is mobility speed; future work will focus on faster overlay convergence and more efficient aerial protocols. Additionally, sharing memory across public infrastructure raises significant security and privacy concerns, necessitating hardware-accelerated encryption. Finally, new programming models are required to help developers manage the inherent volatility and heterogeneity of disaggregated edge resources.

Summary updated

BibTeX
@misc{vaquero2021disaggregated,
  author = {Luis M. Vaquero and Yehia Elkhatib and Félix Cuadrado},
  title = {{Disaggregated Memory at the Edge}},
  year = {2021},
  eprint = {2102.03124},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.03124},
}
Figure