So you have the alert to trigger, then start a 15 minute countdown before the alert actually becomes an alert, it's basically in a stand-by mode for 15 minutes. Within that 15 minutes, if the system falls out of the 90% it's going to reset the alert, then again, the next time it polls, if it goes back above 90, it starts the 15 minute countdown. If you don't need the 15 minute timer, I'd remove that. Unless the CPU is continously spiking up and down, it shouldn't retrigger the alert. The other areas I'd look at are:
- do you have a reset action configured?
- is this a server with multiple CPU
- If you look at the CPU on the node details page, does it hold steady or does in vary significantly
- If multiple CPU are there some at 90 and some below? This would help explain the behavior you are seeing.