[3/3] Google's AI-powered next-generation global network: Built for the Gemini era (Category: Infrastructure)
[3/3] Google's AI-powered next-generation global network: Built for the Gemini era (Rubric #Infrastructure)
Continue. story About the evolution of Google networks, it is worth talking about how new approaches to the architecture of networks solve the challenges voiced in the Internet. past post
1. Exponential scalability The network must flexibly withstand the avalanche of traffic and data growth, especially in regions where AI computing is concentrated. The principle of "WAN is the new LAN" is implemented through the rejection of monolithic in favor of horizontal scaling. (Multi-shard network architecture). Shards are independent – each has its own set of controllers and channels. This allows you to increase the capacity in parallel - 2020 along 2025 Google’s global WAN capacity increased in the year 7 once. In addition, this segmentation simplifies management: each sharding subnet is more controlled in size.
2. Reliability is higher than the traditional “five nines”. The industry usually talks about 99.9percent 99.99% availability, but for critical AI loads, long tail emissions are squeezed (We need determinism and uninterrupted network operation). In practice, the network should localize problems and automatically bypass them before users or processes notice a failure. For that.
- Chards are isolated from each other. (failures do not correlate) In addition, regional isolation has been introduced so that local problems do not cascade globally. Protective ReRoute technology is created to quickly detect communication losses and redirect traffic in seconds After the inclusion of Protective ReRoute, the total downtime for incidents was reduced to 93%.
3. Programmability Driven by Intent (Intent-driven programmability) Google serves billions of users and many corporate customers with different requirements.
- Someone is critical of the delay.
- Someone cares about encryption. Someone has to spread the data geographically. (regulatoryly)
To meet these different requirements, the guys made the network fully programmable. (SDN) high-level policies (intent)That is, created. Unified network representation models (MALT - Multi-Abstraction-Layer Topology) Open APIs for management Centralized SDN controllers that can interpret the intent of operators or applications and apply them to the network. This flexibility allows you to set policies for specific applications or data. (For example, to ensure that a certain type of traffic travels only through nodes in a given country to respect data sovereignty, or that critical services always have backup channels.). High-level management does not require manual configuration. (How in SQL it is enough to specify what you need, and the smart network will adapt to the request)
4. Autonomous network Networks have already passed the way of the form: manual control> automated (script) -> automatic (stringent n). The new goal is to make the network self-managed through machine learning and a digital twin, where models are constantly trained on telemetry. In this way, the network will be able to simulate and anticipate failures, quickly localize the cause of problems, and even optimize channel capacity planning for the future. After the introduction of these tools, the failure response time was reduced from hours to minutes, which significantly improved the efficiency and sustainability of the network without human intervention.
Following these four principles, Google has implemented a number of technological innovations in its next generation network. All this turns its global network into a platform capable of meeting the needs of AI without compromising the user experience. The end of the article emphasizes that such a network opens up opportunities not only for Google, but also for cloud customers. (||A little advertising won’t hurt.||)
In the last post, we compare this article/info from Google and article Meta banned in Russia.
#Software #DevOps #Architecture #Economics #DistributedSystem #SystemDesign