DRS takes the active memory into account of the virtual machine when determining which virtual machine to migrate. It will also look at the stabletime, meaning the time a virtual machine was using the same amount of memory. If a virtual machine memory usage fluctuates heavily, this candidate is seen as less optimal as DRS cannot predict the effect of a migration on the consumption of memory on the destination host, which in turn can affect the exisiting virtual machines.
Therefore determining which virtual machine will be moved by DRS is a challenge. Consumption, stable time of the last 60 minutes, constraints such as affinity rules and that for both CPU and Memory. Depening on which resources is at premimum state within the cluster (which is used more, cpu or memory) DRS decides what metric to weigh more.
When migrating virtual machines for a maintenance mode, DRS considers these migration mandatory. Mandatory moves need to be satisfied before manual moves, and are always executed. DRS will take the overall load balance of the cluster into account and the individual host utilization. It tries to find the best suitable location for this virtual machine.
It's very difficult to calculate the cost of a vMotion, I recently published an article about the metrics involved: http://frankdenneman.nl/2012/12/04/calculating-the-bandwidth-usage-and-duration-of-a-vmotion-process/
Multi-NIC vMotion configurations are highly recommended, as vMotion will load balance single and multiple vMotion operations across the uplinks, providing you with reduced migration times. This in turn can provide DRS the ability to load-balance the cluster in smaller number of invocations.
Im currently publishing a series of articles on Multi-NIC vMotion network design, this might be interesting for you as well: http://frankdenneman.nl/2012/12/18/designing-your-vmotion-network/