Industry Trends in Networking
The long journey to factorize Network Management problems by decoupling policy from configuration has let to a variety of outgrowths of controllers to mange networks.
SDN
During this journey we faced different approaches on Network Programmability Models. We started with vendor specific API or simple CLI/SNMP, Classic OpenFlow based SDN, Hybrid SDN based on “ships in the night” or “integrated mode”, Network Virtualization with Virtual Overlays to finally arrive at “policy intent” or also described as “Managing Networks through abstractions”.
The Promise of OF/SDN had been
“Decoupling Policy from Configuration”
So what actually is SDN:
You can’t just buy SDN.
It’s an architecture which you have to embrace and life
From Cloud over SDN towards Abstractions to a final fully distributed system
2008 - Cloud
In 2008 we started with cloud based offerings. Now in 2016 customers finally move applications or at least parts of mashed application into the cloud.
Core Cloud Infrastructure Services (IAS):
- Compute
- Storage & Content Delivery
- Database
- Networking
Platform Services (PAS):
- Analytics
- Enterprise Applications
- Mobile Services
- Internet of Things
Developer Productivity Services (SAS):
- Developer Tools
- Management Tools
- Security & Identity
- Application Services
In the beginning the primary reason for utilizing cloud resource had been to smooth out peak datacenter workloads also called “bursting into the cloud”.
Later the main motivating factor for companies to move to the cloud is the ability to pay for it as an operating expense, rather than capital expense. Today, base rent and utilities payments for colocation space can be treated as Opex and hardware can be leased and therefore paid as Opex.
Once a company spends more then $200,000 per year for cloud servcies, they should look at the other options. Beside trustworthiness the biggest driver to move back from public to private clouds is simply cost as there are economies of scale once you get bigger. For a lot of bigger enterprises it is simply economically not feasible to stay in the cloud. Shifting resources back into the company may sound strange as most technology companies outsource product and software development to partners and other companies in a bid to save costs and to strategically conserve cash and resources. Problem with scale back outsourcing is also that meanwhile the former staff has moved on so all the knowledge is lost.
2011 - Open Flow Networking
2012 - SDN - Software Defined Networking
2013 - Open Daylight Project
2013 - Network Function Virtualization
2013 – DevOps, The API driven Datacenter
2014 - Managing Networks through abstractions
2016 - Metadata Driven Networking
2018 - Atomic Services
Today’s DC Architectural Battle
System administration is over – we should stop doing it
Web Approach (MSDC)
- IT infrastructure core of its business
- Warehouse Datacenter
- Scale-Out Architecture
- ~100.000 of physical servers
- Single Application Optimization
- Many smaller services
- Application Designed for Failure
- Automate everything possible
- It’s all about being super-cheap commodity systems; costs must grow in a “sub-linear” fashion
- Open Source
- Backbone Bandwidth Calendaring
- TDM style provisioning with custom TCP stack
- L3 Topology
Enterprise Approach (EPDC)
- IT infrastructure is an expense
- “Discovery” Datacenter
- Scale-Up Architecture
- ~10.000 physical servers
- Hundreds to Thousands of Applications
- Application trust boundaries
- HA failover model
- Transactional
- Application specific Infrastructure
- Commercial Of The Shelf
- L2 Topology
Driven by todays applications and digitization efforts the datacenter and in particular the MSDC’s are leading the agenda and there is an ongoing DC Architectural Battle based on misassumptions on requirements.
Main limitation factor for Enterprise to compete with MSDC offerings is typically not the network itself, this are old isolated and not mashed applications. While apps gets re-written to address the digitization needs with modern software development tools the expectation is that the network will be able to be programmed likewise by current development tools which leads towards a DevOps model.
As new application development no longer happens in isolation and the production system is becoming the development platform the logical next question coming up is why can’t I simply use my datacenter tools like PUPPET, CHEF and ANSIBLE likewise for Enterprise networks?
However, main takeaway from that ongoing battle is: “System administration is over – we should stop doing it” and what could be achieved by “Efficiency through abstraction and automation” based on Device to Admin Ratio comparisons over the last years.
Efficiency through abstraction and automation
Device to Admin Ratio
2009
2013
Traditional IT: 50:1
Amazon: 200:1
Google: 10000:1
Traditional IT 50:1
Amazon 10000:1
Google: 30000:1
Conclusions for Enterprise Controllers
From this discussion there are some main design points derived when it comes to requirements for controllers for Enterprise networks.
Primary:
- Reduce Network Complexity
- Brownfield Support – No Software / Hardware upgrade required
- Low Risk adoption of SDN
- Enterprise Scale for real life production network use
- Product with minimal to no programming requirement
- Auto-Translation of high level business intent into network control function
Secondary:
- Abstraction and Automation of Manual Network Operations
- Advanced Visualization (HTML5/Java code with object oriented interface)
- Start with small set of solvable problems (QoS, ACL management , Zero Touch Deployment and IWAN as key applications with identifiable metrics (OPEX savings , ROI)
- Elastic Services Infrastructure ensures scaling as adoption grows
- Advanced analytics for real time network visibility and faster response time
