Google’s Cloud Platform is enabled by its “Andromeda Network Virtualization” platform — a software-defined networking (SDN) framework extending across the entire hardware/software stack. Google takes a holistic approach to all layers in order to achieve QoS, latency parameters and fault tolerance. Google is making huge investments in its cloud data centers — at least $2.9 billion in additional planned data centers worldwide in support of cloud services. As reported in the press, managing the energy consumption of these huge facilities is as important as ongoing capital spending. The company has an extensive CDN footprint to reduce delivery times to end users.
In talks at previous Open Networking Summits, Google first described its global Software-defined WAN, describing the massive backbone as a homegrown project that has enabled the company to treat WAN bandwidth as though it were LAN bandwidth.
Some notes from Amin Vahdat’s presentation:
- Google adopted a share infrastructure model pretty much since its inception. Hundreds of application run across the same cloud platform.
- Google is into cluster networking, where storage is disaggregated into pools for virtual machines
- Google has found that a logically centralized/hierarchical control plane with a peer-to-peer data plane beats out decentralized models.
- The Andromeda Network Virtualization stack runs on this centralized/hierachical control plane model.
- Network virtualization bring challenges: virtual networks must be isolated from each others, DDoS attacks are a threat, network functions virtualization is key to the operation of these networks.
- Looking ahead a few years, based on Amdahl’s Law, Google calculates that future clouds will require extremely high bandwidth and low latencies to interconnect compute resources and shared pools of Flash storage.
- Virtual machine migration capabilities will be key to support high-availability infrastructure.
- Vahdat believes that commodity servers running control protocols will be widely used for SDN.
- Google wishes to eliminate ‘one-off’ protocols and siloed middle boxes in its network.
- Google is looking at programmable packet processors to extend network functionality
- NFV is really about virtualized SDN
- Andromeda is capable of splitting NFV between soft switches and commodity packet processors.
- Andromeda capabilities also include cloud load balancing, virtual machine migration, storage services, bandwidth provisioning, and network configuration tools (ACLs, firewall and isolation)
The full keynote has now been posted online by @OpenNetworking Summit.
