In my experience disk i/o is the biggest bottleneck. It used to be sync()ing writes to disk for strict consistency but that's been pushed down to the DB now. I just looked at my DB systems and CPU is low but disk is nearly pegged.
My data sets are far too big to fit into memory/cache. Disk pressure can be alleviated by optimizing queries but it's a game of whack-a-mole.
I have exhausted EBS i/o and been forced to resort to dirty tricks. With RDS you can just pay more but that only scales to a point – normally the budget.