Networking in the data center: The modern network is flexible

The demands on data centers are currently increasing enormously. New applications in the fields of artificial intelligence (AI), deep learning and high-performance computing (HPC) are emerging around the world. Because companies are using AI in more and more areas. For example, AI systems enable marketing to be personalized or customer support through virtual assistants; they form the basis for recommendation systems and facial recognition. Deep learning applications search for patterns in huge data pools consisting of millions of videos, photos, audio files and documents. In recent years, HPC has helped weather forecasts and flow simulations in aircraft and vehicle construction to achieve previously unimagined precision.

These new tasks can no longer be mastered with conventional applications that are set up on a single computer system. They have been replaced by distributed applications that spread across multiple hardware instances and device classes in the data center. This in turn requires a changed organization of the data centers: Your infrastructure today has to be flexible and adapt to the structure of the applications. CPUs, GPUs specializing in AI, RAM and storage have to connect to one another in ever new configurations and at different performance levels. At the same time, however, this means a significantly increased management effort for those responsible in the data centers. The additional work can be managed through increased automation and the central control of resources in the data center.

But that’s not all: data centers have to be equipped with high-speed networks to connect the various components that are interconnected for a distributed application. Otherwise there is the effect that, for example, the CPUs and GPUs used offer high computing power. However, since the data cannot be transmitted at sufficient speed due to the poor performance of the network connections, the waiting times regularly result in idle phases. CPUs and GPUs could work much more effectively in this scenario, but they always lack data replenishment.