Distributed point-to-point routing method for tasks in cloud control systems

Guan Wang, Yufeng Zhan, Yuanqing Xia, Liping YanJournal of Systems Engineering and Electronics 2022 · DOI

Abstract

With the rapid development of cloud computing and control theory, a new paradigm of networked control systems called cloud control systems (CCS) is proposed to meet the requirements of large-scale and complex applications. Currently, cloud control systems are mainly built by using a centralized architecture. The centralized system is overly dependent on the central control plane and has huge challenges in large-scale heterogeneous node systems. In this paper, we propose a decentralized approach to establish cloud control systems by proposing a distributed point-to-point task routing method. A considerable number of tasks in the system will not rely on the central plane and will be directly routed to the target devices through the point-to-point routing method, which improves the horizontal scalability of the cloud control system. The point-to-point routing method directly gives a unique address to every task, making inter-task communication more efficient in a complex heterogeneous and busy cloud control systems. Finally, we experimentally demonstrate that the distributed point-to-point task routing approach is compatible against the state-of-the-art central systems in large-scale task situations.

Introduction

Traditional Cloud Control Systems (CCS) manage infrastructure by unifying cloud, edge, and terminal devices. However, most current architectures are centralized, relying on a master node to manage worker swarms. As these systems scale, they face a critical central bottleneck where the control plane becomes overwhelmed by scheduling tasks across heterogeneous nodes. Furthermore, standard IPv4 addressing lacks semantic meaning—it identifies location but not capability—and task migration creates significant overhead for central registries.

This paper introduces a decentralized, point-to-point routing method that replaces the central bottleneck with a unified, semantic addressing scheme. By transforming task scheduling into a routing problem, tasks can find their ideal execution devices directly and efficiently.

Proposed Methodology

Unified Semantic Addressing

The core innovation is a semantic addressing scheme that uses a Restricted Boltzmann Machine (RBM) to encode task and device attributes into a "Uniform Code" (UC). Unlike random IDs, this encoding is based on functional characteristics such as CPU requirements and memory limits. Consequently, similar tasks and devices are mapped to proximal locations in the address space.

The Task Addressing Encoding (TAE) process using RBM to generate a Uniform Code (UC) from task and device attributes.

Uniform Address Ring and Distributed Routing

All tasks and devices are mapped onto a logical ring. Scheduling a task is simplified to routing it to the "closest" device group on the ring that matches its requirements. To navigate this ring without a central directory, each node maintains "D-Buckets" (routing tables) inspired by peer-to-peer protocols like Kademlia.

Mapping of tasks and device groups onto a logical Uniform Address Ring for decentralized routing.

To select the optimal device within a group, a Dynamic Quality of Service (DQoS) metric balances performance and reliability. This metric automatically prioritizes reliability in smaller groups where failure is more critical, and performance in larger groups with ample redundancy.

Experimental Evaluation

The system was evaluated using a testbed of 60 virtual machines (8-core AMD EPYC processors, 16GB RAM) comparing the proposed Distributed Point-to-Point Task Routing (DP2PTR) against a standard Kubernetes (v1.19.2) setup. Large-scale overlay network simulations were also conducted using OverSim to model massive environments.

Results and Discussion

Experimental results demonstrate that the Task Addressing Encoding (TAE) method significantly outperforms traditional IP-based lookups as task counts grow, primarily due to the elimination of kernel overhead. When processing $10^5$ task submissions, DP2PTR showed mean elapsed times that were orders of magnitude lower and more stable than Kubernetes, which exhibited exponential growth in response time under heavy loads.

Key insights include:

  • Scheduling as Routing: Functional compatibility ensures tasks "flow" naturally toward compatible devices.
  • Semantic Proximity: Network distance reflects functional capability rather than just latency.
  • Reliability Buffering: The DQoS logic effectively maintains high success rates by adjusting selection priorities based on group size.

Conclusion

While the decentralized point-to-point model offers superior scalability and efficiency, future work will focus on optimizing the RBM weight matrix through machine learning and introducing robust security protocols to protect routing tables in decentralized environments.

Summary updated

Abstract

With the rapid development of cloud computing and control theory, a new paradigm of networked control systems called cloud control systems is proposed to meet the requirements of large-scale and complex applications. Currently, cloud control systems are mainly built by using a centralized architecture. The centralized system is overly dependent on the central control plane and has huge challenges in large-scale heterogeneous node systems. In this paper, we propose a decentralized approach to establish cloud control systems by proposing a distributed point-to-point task routing method. A considerable number of tasks in the system will not rely on the central plane and will be directly routed to the target devices through the point-to-point routing method, which improves the horizontal scalability of the cloud control system. The point-to-point routing method directly gives a unique address to every task, making inter-task communication more efficient in a complex heterogeneous and busy cloud control systems. Finally, we experimentally demonstrate that the distributed point-to-point task routing approach is compatible against the state-of-the-art central systems in large-scale task situations.
BibTeX
@article{wang2022distributed,
  author = {Guan Wang and Yufeng Zhan and Yuanqing Xia and Liping Yan},
  title = {{Distributed point-to-point routing method for tasks in cloud control systems}},
  journal = {Journal of Systems Engineering and Electronics},
  year = {2022},
  doi = {10.23919/jsee.2022.000079},
}
Figure