In constructing its “Jupiter” fifth-generation data centers, Google is essentially the bandwidth equivalent of the Internet under one roof.
Some key takeaways from the presentation:
- Google will open source its gRPC load-balance and app flow-control code
- Google’s B4 software-defined WAN links its global data centers and is bigger than its public-facing network
- Andromeda Network Virtualization continues to advance as a means to slice the physical network into isolated, high-performance components
- Google is deploying its “Jupiter” fifth-generation data center architecture. Traditional designs and data center switches simply cannot keep up and require individual management, so Google decided to build its own gear.
- Three principles in Google’s data center network are: Clos Topologies, Merchant Silicon, and Centralized Control. Everything is designed for scale-out.
- Load balancing is essential to ensure that resources are available and to manage cost.
- Looking forward, a data center network may have 50,000 servers, each with 64 CPU cores, access to PBs of fast Flash storage, and equipped with 100G NICs. This implies the need for a 5 Pb/s network core switch — more than the Internet today!
The #ONS2015 keynote can be seen here:
