I attended the SDN/MPLS conference in Washington, D.C. last week, where I presented on the importance of analytics for WAN SDN application bandwidth scheduling and the need for even richer analytics when looking at the data center, network edge and WAN SDN holistically. In my presentation I highlighted the importance of accurate traffic demand matrices and the need to consider failures when selecting paths, so that the network can survive them without creating congestion. I was not the only one talking about WAN orchestration and analytics.
One of the most interesting presentations in my opinion was by Douglas Freimuth of IBM. Douglas presented his work titled “Orchestrated Bandwidth-on-Demand for Cloud Services.” It is a collaboration between IBM, Ciena, and AT&T. They carried out the work in a laboratory test bed.
In the test bed, there were three data centers (Los Angeles, New York and Chicago) running OpenStack. When VM workload in the Los Angeles data center exceeded a threshold, some of the VMs were moved to the New York data center to reduce the load. Meanwhile, the east-west data center traffic flowed through the WAN. For this the team used a bandwidth-on-demand application and orchestrated an optical path across the WAN. When the Los Angeles data center load returned to normal levels, the VMs were moved back.
What would have happened if there had been insufficient network bandwidth to New York but available bandwidth to Chicago? It was not clear from the presentation if the Chicago data center would have been chosen instead. I am sure that even if this was not the case, it must be in the future plans.
Packet Design’s Network Access Broker analytics can improve Douglas’ work by ensuring that there is sufficient bandwidth even under failures in the WAN. In addition, it does not require a dedicated inter-data center network where all demands are received via the SDN applications.
John Evans and Michael O’Gorman from Cisco Systems presented “Cloud Workload Placement over a Wide Area Network.” Their research attempted to answer where to place individual workloads based on their communication patterns and the availability of data center and network resources. They concluded that the most efficient scenario for most workloads is when both data center and network resources are optimized together. Results showed that optimizing for the data center alone hurt network efficiency and vice versa. This of course requires traffic demand matrices such as those provided by Packet Design.
David Kao and Leon Zhao from Time Warner Cable presented on multi-layer optimization. In multi-layer optimization, ingress-egress router-level IP traffic demand matrices are used to compute the best Layer 3 topology over a physical optical topology. For each demand matrix there is an optimum such topology. Despite this, they concluded that doing no optimization at all gives the best results when the traffic demand matrix is highly dynamic—which is the case in IP networks. The recommendation is to follow the physical optical topology at the IP layer whenever possible and only introduce router bypasses where the expected benefits are undisputable.
To read more about my presentation at the SDN/MPLS conference, visit http://www.packetdesign.com/blog/analytics-orchestration-from-sdn-mpls-conference