[2/3] Google's AI-powered next-generation global network: Built for the Gemini era (Category: Infrastructure)

[2/3] Google's AI-powered next-generation global network: Built for the Gemini era (Rubric #Infrastructure)

Continue. story As for the evolution of Google’s networks, it’s worth noting that they are now seeing a new turning point – the explosive development of artificial intelligence, which places unprecedented demands on the network. (For example, training large models dramatically change the load profile on the network.). In fact, there are as many as four separate calls that lead to a change in the design principles of network deployment.

1. WAN as the new LAN Learning modern foundation models requires combining thousands of TPUs/GPUs. What used to be within a single data center is now geographically distributed. (The continent is similar to the data center. message banned in Russia Meta). The network needs to scale an order of magnitude larger to link remote clusters as if they were in the same local segment. At the same time, traffic from distributed learning comes in bursts that need to be effectively detected and routed without losing performance.

2. Zero tolerance for failure AI model learning processes and large-scale inference are very sensitive to disruption. Stopping training due to network failure is unacceptable due to downtime of expensive iron. The network is now expected to 100The network must be designed so that any fault-tolerant mechanisms are triggered instantly and do not affect the long learning process.

3. Increased security and control requirements The data on which models are trained and the models themselves are a valuable and sensitive resource. They need to be protected from both leaks and unauthorized changes. In addition, as AI spreads, demands for compliance with regional regulations and for data control on the fly are growing. (transit). The network must ensure isolation, encryption, compliance with the policies of different countries and companies, so that AI services remain reliable and law-abiding.

4. Operational excellence with increased complexity Scale growing by an order of magnitude cannot be managed in the old way. Google is applying SRE best practices and is already using machine learning to manage the network, but now the goal is to minimize the human factor. The network must operate with a minimum of manual intervention, because linear build-up of infrastructure would otherwise result in unmanageable increases in complexity and cost. New approaches are needed to automate, quickly identify and resolve problems, and optimize capacity.

Hence the new design principles of networks, which we will discuss in next.

#Software #DevOps #Architecture #Economics #DistributedSystem #SystemDesign