Amin Vahdat (Distringuished Engineer and Tech lead for networking at Google) has been a key contributor to Google’s SDN efforts. Amin took the stage at #ONS2014 today in the opening keynote to speak about Andromeda – Google’s internal codename for their infrastructure services. A thorough and to the point keynote had the audience rapt and Amin’s knowledge and passion for network virtualization came through in a very effective presentation.
What people used to think of the cloud was – it’s a way to get on demand access to compute. The little understood appeal of the cloud is a fundamentally easier operational model with much higher uptimes and efficiencies. State of the art infrastructure services such as preventing DoS attacks – providing load balancing and storage-on-demand are much more important today. Cloud and SDN/NFV provide programming models unavailable elsewhere combined with scalability combined with low latency programming at massive iops.
Andromeda network virtualization
Andromeda is Google’s internal codeword for network virtualization in their internal networks – providing SDN control of the entire hardware and software stack – it covers QoS along with low latency and fault tolerance. Virtualizing the SDN means creating extensible network function virtualization to orchestrate and manage the network provisioning high availability infrastructure.
Andromeda’s physical network is similar to others – it has racks with top-of-rack routers aggregating the traffic; and the logical networks running over it with their own IP addresses and NFV with load balancing, DoS prevention, Access Control Lists etc.
Google eliminates latency by having a good CDN with low latency built in – they are creating what’s being referred to as “cluster networking”: building huge switches out of cost efficient building blocks combined with the disaggregation of storage – with more efficient scheduling. The aim is to run computation when and where needed and not dependant on locality. Google has been running shared infrastructure since it’s inception – and it has been the basis for many commonly used scalable open-source technologies (GFS as the basis for HDFS in Hadoop, BigTable, etc.
Challenges that come with Network Virtualization
New challenges emerge from implementing network virtualization – such as isolation, distributed-denial of service attacks,creating virtual IP networks, NFV, mapping external services into internal namespaces, authentication, authorization and billing. And the most important of it all is maintaining efficiency while doing all of the above.
Whither SDN at Google?
A simple definition of SDN at Google: split the control plane from the data plane to allow an independent evolution of data path from control path, with commodity servers running control protocols, allowing for provisioning of an isolated high performance network across:
NICs
Soft switches
Storage
Packet processors
Fabric switches
Top-of-Rack switches etc
As an example, the Andromeda control stack controls VMs spread across the network fabric associated with the Top-of-Rack switch. Some VMs are also associated with soft switches, Fabric switches, Packet processors, cluster routers and, storage. The Andromeda controller controls all of them – the functionality of all the software and hardware to provide end to end QoS and performance.
Case studies
Two case studies were presented – the first was the Andromeda network datapth with integrated programmable nfv – the datapath pipelined and critical optimizations applied end-to-end with the goal of obtaining near native performance and cpu efficiency.
The results: Datapath throughput went up 4 to 6 times improvement vs the baseline throughput! And that’s before another 2-3 times more optimization that is still possible (for a total 8 to 16 improvement that’s possible to achieve) on the same machine.
A second study showed similar optimizations with Google Cloud Load Balancing – provisioned in under 5 minutes taking 4 seconds to ramp and reaching steady state in under 120 seconds – all for a cost of about $10 on the Google cloud. Impressive it does sound – but it would have been nice to hear what exactly was being load balanced so it could have been put into perspective.
Google has been leveraging their decade plus of experience in delivering a high performance shared computing infrastructure for a logically centralized SDN control to orchestrate VMs NICs and fabrics etc. The goals of near native performance with a scalable and reliable architecture for a flexible and extensible NFV seem to have been met.
Now if Google would only make some of these impressive technologies Open Source – that would make things really interesting in the SDN and NFV space!
This is Kshitij Kumar (aka KK) signing off from the Keynotes on the concluding day of #ONS2014 – follow me at @kshitijkum or feel free to reach out to me at the event!
原文：
http://opennetsummit.org/blog/2014/03/andromeda-internal-sdn-and-nfv-at-google/
