The Path of an AI Creator News: Announced at Cisco Live! AI will dramatically transform networks. How will developers change? #AI #Network #CiscoLive
Video explanation
Will AI revolutionize the world of networks? Just like the Internet in the 90s!
Hello, this is John, your familiar voice in AI technology commentary. Today, I'd like to talk about how AI is going to bring about incredible changes to the world of "networks" that we are familiar with. It may have an impact as big as when the internet appeared in the 1990s and completely changed the world!
Chuck Robbins, CEO of Cisco, a major network equipment company, also described the impact of AI on networks as "a return to the Internet of the 90s." At the time, the advent of TCP/IP (the basic protocol for Internet communication) broke down the walls of closed computer networks and connected information around the world.
Advances in AI are creating entirely new demands on networks:
- Programmability: The network can be flexibly controlled by programs.
- Observability: The ability to visualize and understand the network status in detail
- Optimization: "Optimizing" the network to make it run more efficiently
At first, this may sound like something only network engineers would talk about. But looking back at history, any major change in infrastructure (the foundation that supports society and services) will eventually affect the work of software developers. So, developers, please keep reading and think, "This is not just something that concerns me!"
Goodbye, just a box? Network devices are getting smarter!
In the world of "cloud native" (a design philosophy that makes the most of the benefits of the cloud), which we often hear about these days, containers (a collection of components for running an application) and APIs (application programming interfaces (APIs): a gateway for software to exchange information with each other) have become the common common language. In fact, the world of networks is also in the midst of a similar major change.
According to Thomas Graf, CTO (Chief Technology Officer) at Cisco (he is also the developer of a famous open source technology called Cilium), network devices such as routers and switches, which have been "just boxes" until now, willProgrammable and intelligent platformIt seems that it will change to.
For example, new switches equipped with DPUs (data processing units: smart chips installed in network devices that specialize in data processing) have appeared. These are amazing! In addition to traditional switch functions, firewalls (which act as a barrier to prevent unauthorized access) and load balancing (a function that organizes work traffic and distributes the load) are built into the switch itself. In other words, functions that previously required separate dedicated devices are now integrated into the network system itself.
Graf calls this DPU-equipped switch a "programmable system." This makes it possible for the switch itself to provide network services that previously ran on virtual machines (virtual computers created within a computer) or other containers. Furthermore, by combining it with mechanisms such as eBPF (a technology that safely and efficiently runs programs in the kernel, the heart of the OS), traditional network operations such as firewalls, network partitioning, and status monitoring can be managed more flexibly with code.
Graf says that the style of network operations will also change from the previous "ticket management" (a method of issuing a request form = ticket every time a change is requested) to "GitOps" (a new method of managing network settings with a system called Git, in the same way that program code is managed).
Can AI help with network problems? Here's how debugging will change in the future!
Cisco recently announced an AI model specifically for security measures, as well as a conversational AI assistant called "AI Canvas" specifically for network operations. It's so futuristic!
According to David Zacks, director of AI and machine learning at Cisco, "AI systems are not smarter than network engineers.AI has access to much more data than humans" I see! A large amount of telemetry (data on the operation status and performance of the equipment) is collected from the network devices, and the AI analyzes it (called machine inference) and presents specific countermeasures. This is becoming the "standard" technology for operating a stable network.
As this cycle of analysis and improvement using AI progresses, software developers may one day be able to use this mechanism in the development stage. For example, it may become possible to simulate in advance how an AI-based application will actually work on a network, and automatically find areas where performance suddenly drops or processing blockages (bottlenecks). This looks set to greatly increase development efficiency!
A new network design for AI? The boundary between apps and infrastructure disappears!
One of the themes that was repeatedly mentioned at this Cisco event was that "in order to support AI, we need to rethink everything from applications to infrastructure, and as a result,The line between applications and infrastructure is becoming blurred"about it.
According to Cisco's Jeet Patel, "Both AI models (like the brains of AI) and the semiconductors (silicon) that run them are becoming more and more custom-made." AI models are becoming smaller, semiconductors are becoming easier to program, and development times are becoming shorter.The AI model itself becomes part of the applicationIn other words, improving the app is almost the same as improving the AI model.
In this way, the close proximity of app logic (processing flow) and AI inference hardware (the machine that AI uses to think) has made it necessary to fundamentally rethink the entire design philosophy (architecture). In order for AI-based processing (workload) to run smoothly, developers also need to understand how the design of the AI model, the network communication speed (bandwidth), and where the AI inference processing is performed (inference placement) affect each other.
In particular, LLMs (large-scale language models: AI trained on large amounts of text data) like ChatGPT exchange large amounts of data, so they are very sensitive to communication delays (latency) and congestion. These types of problems are often difficult to see until the network is actually under strain.
In future AI development, it will be important to think in terms of directly mapping the flow of AI processing onto the network structure (topology). For example, strategies being discussed include distributing AI work across multiple routes (pipeline parallel processing), placing AI inference processing in places with good network conditions, and storing parts of AI models near where calculation processing is performed in advance (caching).
This is infrastructure design at the developer level. We are now in an era where the performance of AI applications is determined not only by the performance of the GPU (a semiconductor that is good at AI calculations), but also by "where," "how," and "how fast" the data flows.
Will developers finally have control over the network?
So, is the day approaching when software developers will be able to directly programmatically control the network?
According to Jim Frey of the Enterprise Strategy Group, "Network programmability (the ability to control it by program) has been a long-standing goal in the networking industry." There is even a term called "NetDevOps (network DevOps: the idea of automating network operations like development)," and there is an active community working to realize this.
However, Frey said, "Achieving this goal was extremely difficult because there were no standard interfaces (common standards for connection) and each device manufacturer had their own closed design."
However,The advent of AI is changing the rules of the game.This new world is forcing network infrastructure teams and equipment manufacturers alike to rethink how they do things and find ways to bring programmability into line with other infrastructure disciplines.
Given this new reality, Cisco believes that it is not a pipe dream to think that in the future, AI developers will be able to control the network (declarative access) by simply declaring, "Please provide this level of communication speed and low latency for this app!" and the network will automatically respond to that request.
Cisco's Patrick Le Maitre said, "We are not just treating AI as a workload.It is being built as a platform"That will change everything," he said. AI will play a key role and even change the nature of networks. That future may be just around the corner.
A word from John
Wow, AI really does have an impact on many different fields! This time we talked about networks, which are extremely important infrastructure that we don't usually think about. But isn't it exciting to see how AI is trying to bring about change at the foundational level of society? I hope that this will lead to a more convenient and comfortable future not only for developers, but also for us users.
This article is based on the following original articles and is summarized from the author's perspective:
Cisco Live: AI will bring developer workflow closer to the
network