IT stresses created by the business and technology trends described above have catalyzed several significant changes in how data center networks and systems are designed, implemented and managed. We’ll detail the infrastructure trends and evolution below, but in sum, these include:
- Design patterns that improve scalability and performance such as n-tier CLOS fabrics.
- The pervasive use of virtualization throughout the stack such as VMs, containers, NV and NFV.
- The use of SDN, programmable infrastructure and automation.
- Aggressive data collection to feed infrastructure analytics for performance optimization and security with use of intent-based systems, AI and machine learning.
- The creation of hybrid cloud infrastructure and services that integrates public clouds across the XaaS spectrum into IT operations.
- An endorsement and adoption of open hardware and software based on interchangeable commodity equipment, freely modified software and published APIs.
Together, these serve to evolve data centers and networks built for an era of client-server applications and deployed behind the walled garden of an enterprise network into infrastructure ready for cloud services, mobile users, disaggregated applications, rapid product and software development cycles, massive quantities of data and new forms of AI including deep learning.
Network scalability and performance
The nexus of trends detailed in section one has created unprecedented stress on data center networks, particularly backbones used for east-west traffic. Older, deeply hierarchical network topologies such as fat trees create bottlenecks at the network core that hinder traffic flow between systems such as multi-rack container or Hadoop clusters. For next generation networks, a better alternative is a collapsed, scalable CLOS fabric such as a leaf-spine or collapsed-spline topology. When supporting massive farms of nearly identical servers, such as container, Hadoop or machine learning clusters, a flat topology might be combined with a more traditional routed core into core-and-pod design.