Coordination of Networks in Computing Centric Network

In the future, AI applications would become more popular, and service scheduling would become generalized, even providing a common invoking function interface. At the same time, computing resource would be more ubiquitous, and clients can get access to a nearby computing node to obtain a faster service response speed. In this situation, the coordination of networks is needed, so that computing nodes in different networks can flexibly communicate to each other and complete the client's computing service. In this document, we introduce a preliminary mechanism to coordinate different networks in the Computing Centric Network (CCN).

The CCN in this document means an overlay network or called a federation that focuses on the computing service providing. In CCN, many service nodes in different networks can join in or leave the federation dynamically. Currently, we suggest a distributed way to manage the federation, and thus the coordination of the different service nodes is a key issue in CCN. From the aspect of the client, it does not care where the service is, and the main concern is that the service could be completed in time. In CCN, we assume that the service can be provided either in the clouds or in the MECs. As an example, the federation can own a common anycast address. The associated anycast address will allow clients to access a nearby network service node, however it may not be optimal. For example, in the scenarios of CAN , the nearest node may be busy and can not complete the job quickly because the computing ability in an MEC is limited. In this document, we suggest that it can send this request to other service nodes if it is busy, in order to do the service steering or load balancing. In this situation, the request packet of the client appears like a DNS query, and the network federation will return a proper service node to the client.

In this distributed system, each network service node needs to know the statuses of the surrounding service nodes, such as whether they can accept more client sessions for a specific service, so as to make a better offload. In this way, each node maintains part of the information of the whole federation, but they can provide a resolution mechanism and return a computing node address for the client. A general procedure is described as follows. Firstly, each computing node needs to know the information of some of the surrounding nodes. Secondly, after receiving the client's request, the computing node will decide whether to provide the service by itself or offload to other nodes according to its own status and the computing information of nearby service nodes. The looking up of the proper service node could be recursive. Finally, a feedback is given to the client. As a conclusion, a computing node needs to store and update the computing statuses of neighboring computing nodes that are currently available for the service. If necessary, it can select a proper neighbor node for offloading according to a certain policy.

TBD.