If you have an alarm state set for 95%, the alarm state will be activated if that threshold is breached when the value is polled by the agent.
A spike occuring between polled values will not be reflected.A previously detected spike will appear as a closed, but unacknowledged alarm.
You can set alarms on the 5min load to obtain an alarm state based on the 5min average.
Some other things you could do -> Have your 95% alarm launch a script that does a deeper dive chewing a few more % CPU to gather per second data for the next polling interval.Maybe grab some dtrace output if you are looking to debug a specific operation.
The average CPU Utilization, unlike load, is not available from the kernel, so some script work is required.
Primealert ScriptRunner can be used to pop return values from scripts back up into SunMC for alarming and/or use with PRM/PrimeAlert Reporter.You could craft a script that returned the average of last X values pulled from sar -u.
iMac