A brief discussion on routing and scheduling technology for computing-network integration

1. Conflict between network capacity scheduling and computing capacity scheduling

Traditional services, including public cloud, private cloud, edge cloud or data center, all provide computing power services based on DNS domain name resolution or IP addressing. In short, they only have simple network power scheduling, which is independent of computing power scheduling. The characteristics of this service mode are as follows:

1️⃣ The computing power service status and network reachability status are separated from each other. The network only ensures IP reachability, but cannot ensure the availability of computing power services or the optimal quality of computing power services. There are two main reasons:

A. The optimal network forwarding path is not equivalent to the optimal computing power service provided;

B. The optimal computing power node provided is not equivalent to the optimal network forwarding path.

2️⃣ DNS domain name resolution or IP addressing can only dispatch computing power requests to a unique service node. A single edge computing node usually has limited computing power and cannot link multiple edge nodes, making it difficult to provide fine-grained and elastic computing power services.

The following is a simple example to explain the above characteristics.

Regarding the first question: Consider the network topology composed of routing devices and edge computing nodes as shown in Figure 1 below. The user's computing request data is usually forwarded following the shortest IP path of the router device. However, when the computing request data reaches router W, the edge computing node (MEC 2) is overloaded or even down, then the computing request will be discarded by the edge computing node, resulting in the user's computing request not being responded to in a timely manner. Until the service provider re-issues the DNS update or IP update and redirects it to other computing nodes (MEC 1), the user's computing request can return to normal.

Figure 1 Example of the coordination problem between traditional network and computing power services 1

Even if router W senses that the computing node (MEC 2) is abnormal (usually it can only sense the abnormal link connection status, but it is difficult to sense the load situation), it may use fast rerouting technology to address another backup path and redirect it to the computing node (MEC 1). As shown in Figure 2, from router U to router W, and then to router V, but this temporary redirection path usually cannot guarantee QoS quality (for example, latency, bandwidth, etc. requirements), and it is difficult to provide users with the best computing power service.

Figure 2 Example 2 of the coordination problem between traditional network and computing power services

Regarding the second question, the computing power of a single edge node is usually very limited (for example, the processing power of 10 servers). The processing power and reliability of edge computing nodes are far lower than those of large data centers. It is usually necessary to provide computing services across multiple edge nodes to ensure better service performance and reliability. A key issue is how to direct computing request data to technical nodes that can be used to provide optimal services. The current common practice is that the service provider provides the corresponding computing node domain name or IP, and the user determines a computing request node through a certain hash mapping mechanism and initiates a computing request to a single computing node domain name or IP. Once the computing node is determined, for a single user, its computing request is usually scheduled to a uniquely determined computing node. This method can, to a certain extent, re-switch the computing node through DNS domain name updates or IP updates, but the switching granularity is large, real-time performance cannot be guaranteed, and the user experience is greatly reduced.

In summary, the existing network service provision methods can be summarized into the following two contradictions :

• The contradiction between network capacity scheduling and computing capacity scheduling being isolated and independent from each other;
• The contradiction between distributed network power scheduling and centralized computing power scheduling.

2. Computing and Network Integration Scheduling Technology

IP network routing and forwarding are all based on the distributed hash table DHT (Distributed Hash Table) mechanism. The network and computing power topology in Figure 1 are abstracted to form the following topology:

Figure 3 Abstract topology of traditional network and computing service

Traditional routing technology collects link status information of routing device nodes to form a network topology Graph (V, E) formed by nodes u, v, and w, which consists of node set V and edge set E respectively, and calculates and generates routing information based on this. Computing service nodes X and Y are not included and do not participate in routing calculations.

Further abstract the computing power service nodes X and Y in Figure 3, add them to the routing metric, expand them to the routing calculation domain, form a computing power-aware router, participate in the routing calculation, and form the following network topology:

Figure 4: Computing network topology model

At the same time, the computing power of the computing router (load, processing power, business type, etc.) can be added to the routing calculation as a link cost metric (c4, c5 in the example). Traditionally, the IP routing domain performs routing at the expense of bandwidth, latency, etc.; through computing power-aware routing, computing power such as throughput, computing latency, and server load can be treated as a virtual link state and added to the routing calculation as a cost metric.

The following example forms a computing network routing domain that integrates network overhead and computing power overhead.

Figure 5 Example of computing power-aware network topology

Taking u as the root node, according to the network link topology information, the following path and cost data can be formed. Network cost and computing power cost together form the final computing power service cost.

Figure 6: Computing power-aware computing routing table

The calculation results show that the cost of u obtaining the optimal computing service is 60, and the path is: u->w->v->X. If the calculation is based solely on network cost, the optimal path is u->w->Y. Since the computing power provided by the Y node is limited (it may be due to factors such as excessive load and performance limitations, which are reflected by the value of computing power cost).

Further abstraction, based on existing traditional routing technologies such as OSPF, IS-IS, and BGP, introduces computing power overhead measurement, and uses computing power measurement as leaf information to participate in algorithm calculation, forming a general computing power-aware routing framework, as shown in Figure 7:

1. The computing power node runs in the form of a computing power container (which can be a physical device, server, virtual machine, container, etc.), and the routing process application runs at the bottom.

2. The computing power node is connected to the edge computing node egress router through a Virtual Link or a physical link, and the computing power function type and computing power quality overhead measurement information of the computing power node are published through the routing protocol and flooded to the entire routing domain.

3. The computing power information related to the computing power nodes of the entire network is flooded throughout the network through the routing protocol. After each device node collects the link topology information, it performs computing power routing calculations. Based on the traditional routing table, it calculates and generates a computing power routing table to guide computing power routing.

Figure 7 Computing power aware routing framework

Traditional routing devices are all based on IP routing. In existing network applications, it is still necessary to use IP addresses or address segments to identify computing power services based on compatibility considerations. Typically, this can be combined with AnyCast technology to divide different types of computing power into different anycast address segments, allowing the computing network to be seamlessly integrated. It naturally has the ability to defend against DDos attacks, improve network throughput, reduce network congestion, and have high availability. The availability of the entire network is the availability of computing power, avoiding single point failures.

Computing-network integration involves all aspects, including computing power addressing, addressing, routing, ecology, etc. Different from the computing power routing technology based on SRv6, APN6+ and other technologies at the OverLay level, the above is a brief understanding of computing-network integration technology from another dimension, from the traditional UnderLay routing scheduling dimension.

3. Outlook

The trend of computing network becoming the "main battlefield" of operators has become established. Home edge cloud and home services are the key points for the implementation of new computing network infrastructure. Deep integration of computing network and provision of "one-point access, ready-to-use" computing services will have a revolutionary impact on home services.