A novel CFN-Watchdog protocol for edge computing

Highlights

• Propose CFN-Watchdog, a fault-detection protocol of tasks and VMs for CFN.

• Develop a theoretical model to analyze the performance of the proposed CFN-Watchdog.

• Run extensive simulations that verify the CFN-Watchdog and the theoretical model.

Abstract

Compute first networking (CFN) is a latest distributed framework that intelligently allocates computing resources for edge computing according to computing load and network status. It requires real-time visibility of available statuses of local or remote computing resources. To the best of our knowledge, this paper is the first to propose a centralized fault-detection protocol called CFN-Watchdog to well meet this CFN requirement and timely recycle resources occupied by faults. We then theoretically analyze the impact of various parameters (e.g., detection thresholds, task processing time, and network delay) on the Watchdog performance. Extensive simulations verify the effectiveness of our proposed protocol and the accuracy of our theoretical model. This study is very helpful to optimize parameter configurations and better design fault-detection protocols for edge computing.

Introduction

Edge computing, an extension of cloud computing, is a popular computing paradigm that brings computation closer to users, aiming at providing convenient and fast computing services. It adopts a cloud center for computing-resources allocation. As shown in Fig. 1(a), whenever a task arrives at one edge node, this edge node will send a resource request to a cloud center. The cloud center then determines which type of resources (cloud or edge) should be allocated to the task and send resource allocation responses to the involved edge nodes. Upon receiving the allocation response, the requesting node offloads the task to the target edge nodes. However, this centralized allocation mechanism might lead to a long duration from launching to starting processing a request, due to the network delay between edge and cloud. As a result, it actually goes against the design objective (e.g., fast response) of edge computing.

Recently, compute first networking (CFN) [1], as shown in Fig. 1(b), has been proposed to address the above issue in edge computing. CFN is a novel distributed framework for computing resource allocation and it can dynamically route tasks to optimal computing nodes according to computing load and network status. CFN equally treats edge nodes and cloud nodes, and integrates them into a distributed computing-resource pool. It then introduces a new control plane to manage the pool so as to intelligently allocate resources to meet users’ requirements. Roughly speaking, jointly taking into account of network conditions and available computing resources, the control plane intelligently allocates proximity resources to a user if the user’s request is a real-time task; otherwise, it offloads the task to cloud for processing.

CFN adopts a distributed framework to allocate computing resources for task processing. It constructs a common computing resource pool over a number of geographically dispersed edge/cloud nodes with heterogeneous computing capacities. In CFN, a task may be processed locally or remotely in multiple nodes. The diversity of network connections among these nodes and hence their network delays, as well as the large divergence of nodes’ computing capacities, make it very hard to monitor the status of task processing. Even we can monitor the status, it often takes a long time to detect a task or virtual machine (VM) faults in a distributed framework [2], [3], [4]. Here, a fault occurs when the program of a task or a VM throws a runtime exception. On the other hand, CFN requires that the available status of local or remote computing resources should be visible in real time, in order to efficiently allocate resources from the computing pool. Therefore, we need a mechanism that can detect task or VM faults timely and remotely, so as to quickly recycle resources occupied by these faults.

However, CFN is now in early infancy and does not provide a mechanism to acquire real-time availability of local or remote computing resources. Popular fault detection methods in IP networks or cloud computing might not be applicable for CFN as well. For example, the bidirectional forwarding detection protocol [5] in IP networks is mainly used to detect the fault of network devices, instead of task or VM faults; Hadoop’s JobTracker [6] in cloud computing is mainly used to detect local task faults, instead of remote detection. Therefore, under the distributed framework of CFN where tasks are processed locally or remotely in multiple nodes, it calls for a novel remote fault-detection design to monitor the statuses of task and VM timely.

This paper designs and analyzes a novel fault-detection protocol called CFN-Watchdog (for short, Watchdog), referring to the idea of watchdog timer [7], to address the above visibility problem of local or remote computing resources in CFN. Our contributions are summarized as follows.

• Propose Watchdog, the first centralized fault-detection protocol of tasks and VMs for CFN, that remotely monitors available statuses of distributed computing resources in edge computing. In our protocol, a number of Watchdog clients periodically report the statuses of their monitored VMs to one Watchdog server that is connected to the CFN. Benefiting from the fast response of the centralized protocol, the control plane of CFN can make its managed resource statuses visible in real-time and timely recycle resources occupied by faults.

• Develop a theoretical model to analyze the performance of the proposed Watchdog protocol. Our model factors in important parameters including detection thresholds, task processing time, and network delay and characterizes the false alarm and miss detection of fault events on the system throughput. With our model, we can choose optimal parameter settings.

• Run extensive simulations that verify the effectiveness of the proposed design and the accuracy of our theoretical model.

The rest of this paper is organized as follows. Section 2 summarizes related works. Section 3 outlines CFN and designs the proposed CFN-Watchdog protocol. Section 4 theoretically analyzes the throughput of CFN. Section 5 evaluates the system performance. Section 6 concludes the paper. In addition, Table 1 lists the main notations and their meanings.