Shared todayAI computing power seriesIn-depth Research Report:AI Computing Power Topic: Progress and Prospect of Computing Power Routing (CATS) Technical Standards
Report produced by China Mobile Research Institute).
Report total: 25 pages.
Featured Report**: The School of Artificial Intelligence
In the past two years, China Mobile has continued to innovate and make every effort to promote the development of computing power network, and has achieved a wide consensus in the industry, and the concept is deeply rooted in the hearts of the people. The computing power network is a new type of information infrastructure that takes computing as the center and the network as the foundation, and deeply integrates network, cloud, data, intelligence, security, edge, end, and chain (abcdnets) to provide integrated services.
To meet the extreme needs of new computing power applications, we build an innovative technology system that integrates computing and network to help realize high-performance basic networks. Break through the Internet architecture protocol, superimpose computing power vectors on distance vectors, and optimize computing network resources as a whole. Break through the bottleneck of wide-area transmission performance, realize efficient RDMA transmission and remote transmission, and realize high-performance interconnection between computing. Compatible with Ethernet to the greatest extent, a new type of intelligent computing center network with no blocking, high bandwidth, and low latency is built. Break the boundaries of computing networks, generate endogenous computing power in the network, and realize distributed collaboration of computing tasks across network nodes.
In 2018, he began to study the computing-network convergence technology, and proposed to introduce computing information into the routing domain for joint scheduling in the face of problems such as "performance reversal" of cloud-edge collaboration and edge-edge collaboration. It is found that the introduction of computing power information into the route has a certain optimization effect in the case of low, medium and heavy load. When the load reaches 60%, the available computing power capacity of the overall system is increased by 3317%, and the end-to-end average latency is increased by 3529%。
Computing power routing needs to solve the problems of "what to transmit", "how to transmit" and "how to use", and it is necessary to design an adaptive computing power notification mechanism and a new multi-factor calculation algorithm to realize the joint optimization of computing power and network. There are many dimensions of computing power information, so you need to define high-availability computing information for route scheduling, taking into account the cost of packet packaging and availability. The higher the advertisement frequency, the more real-time the computing power information, but the greater the overhead, how to find the balance between the advertisement signaling overhead and the real-time information. The superposition of computing power vector on the BGP distance vector changes the BGP routing method and affects the BGP routing decision. Simple overlays will cause routes to not converge.
Report total: 25 pages.
Featured Report**: The School of Artificial Intelligence