SDN Controllers

[nivoberg]

How many sdn controllers should be present in a wireless sensor network environment? Is it related to field size or sensor numbers? 

[ShaunWackerly]

Hi nivoberg,

I think that the number of SDN controllers would depend upon the types of data that the wireless sensors would be sending, and how many wireless sensors would be involved. In general when dealing with OpenFlow, it's the total number of PACKET_IN messages that the controller will need to handle that determines when additional controllers are needed. In an optimal use case with sufficient resources, our SDN controller is able to handle 2 million packets per second.

Could you post a little more information about your use case?

Shaun

[nivoberg]

Hello Shaun

Thanks for your reply.

Actually, what you have stated regarding the packet-in messages that the controller can handle is correct. However, from reliability and scalability point of view, experts state that the number of controllers should be odd number ( ofcourse not one since it will face one point of failure) so at least three should exist. 

My simulation will be over a field of 100x100 m2 with number of sensors ranging from 500 sensors to 1500 sensors that will monitor the temperature change and randomly deployed. 

So do you think it is possible to formulate an equation for number of sdn controllers depending on number of sensors, number of packets to be sent, type of connectivity ( fully connected of K-connected ) network? 

Thanks

[ShaunWackerly · Accepted solution]

Hi nivoberg,

The reason for recommending an odd number of controllers is to handle failure scenarios, not for scalability (unless I'm mistaken). The failure scenarios are typically either a control-plane outage (network failure) or the failure of a specific controller. Having an odd number of controllers makes these failures easier to handle for various reasons. For scalability, something like 4 controllers will have better scalability than 3 (despite 4 being even).

Regarding an equation for the number of controllers needed, I think you'll need an estimate of the average packet rate from a single sensor. I'm not sure that the type of connectivity matters, unless that will increase the number of packets sent from any given sensor (say, by increasing the number of hops). I think you could use:

average total packet rate = (number of sensors) x (average pkt rate per sensor)

number of controllers needed = (average total packet rate) / (maximum packet rate per controller)

If this is dynamically computed, you'd need to make sure the number of controllers is never less than the minimum required for reliability purposes.

[nivoberg]

Thank you Shaun