Our architecture uses a central controller (the DC) to manage the DAPs. Each DAPs sends periodic reports to the DC. This raises scalability concerns. To address these concerns, we note that our DC was able to easily manage a network of 24 DAPs and 24 clients, without any special optimizations. The CPU load on the DC never exceeded 30%. We estimate that the amount of control traffic generated by each DAP was less than 20Kbps. Thus, we estimate that a slightly more powerful DC could easily handle a network of about 100 DAPs, without any special optimizations. This should be enough to cover a floor of our office building.
We note here that it is not strictly necessary to use a single central controller. What is necessary is the use of global knowledge while making association and channel assignment decisions. In theory, the functionality of the central controller can be either be replicated, or even implemented in a fully distributed manner. The DAPs can exchange information with each other to gain a global view of the network, and make appropriate decisions. However, this approach is more complex to implement, and has its own set of scalability concerns.
Another issue we must address is the impact of several DAPs in close vicinity, beaconing and sending probe packets. Our measurements show that in the common case, the impact on performance is less than 1%. This is due to two reasons. First, only active DAPs send beacons, and second, when we use multiple channels, the number of DAPs on any one channel is small.
NSDI-2008
