Firstly, allow me to clarify a few points. As you pointed out, spot instances are not meant to be run 24/7 - they are intended to provide extra compute capacity for a short time-span at a lower cost. Essentially, they are intended for tasks that can be broken into small pieces - so a terminated instance will not have a significant impact on the overall task.
That said, I have previously run an instance using the spot request model - its usual uptime was about 3 months - and I would typically overbid by about 15x the average market price. While this used to be a reasonably cost effective approach, I found that as more people employed the same technique, the price volatility increased to the point that it was no longer advantageous over a reserved instance.
The average spot price tends to hover around the hourly cost of a medium utilization reserved instance. Doing the math based on the current spot prices vs heavy utilization instances, you get the following:
+-------------+-----------------------------+------------------------------+----------+
| | 1 year | 3 year | Spot |
+-------------+-----------------------------+------------------------------+----------+
| Small | $0.016/h + $195 = $0.0383/h | $0.013/h +$300 = $0.0244/h | $0.027/h |
| Medium | $0.032/h + $390 = $0.0765/h | $0.026/h+$600 = $0.0488/h | $0.038/h |
| Large | $0.064/h + $780 = $0.153/h | $0.052/h + $1200 = $0.0977/h | $0.108/h |
| Extra Large | $0.128/h + $1560 = $0.306/h | $0.104/h + $2400 = $0.195/h | $0.216/h |
+-------------+-----------------------------+------------------------------+----------+
It is evident that in many cases, the 3 year heavy utilization reserved instance works out to less than the current spot price (which is subject to volatility). From my experience, the actual average spot price tends to be at least 50% higher than the baseline market value, since significant spikes are not uncommon.
Now, to actually try and answer your question:
You are correct that spot instances cannot be stopped - in some ways that does agree with the entire premise behind spot instances. Traditionally, an instance will create the EBS volume(s) it is using from snapshots, but it is possible to script the process of attaching an EBS volume. I used to use this script to setup 2 EBS volumes in a RAID0 (I am pretty sure I modified it from some script I found, but I can't find the original at the moment):
#! /bin/sh
#
# /etc/init.d/mountec2vol
#
# chkconfig: 234 20 50
# description: Assigns an EC2 EBS Volume to a device and mounts the device
#
# To add this as a service run:
# /sbin/chkconfig --add mountec2vol
#
# VARS
#
VOL1="vol-xxxxxxxa"
VOL2="vol-xxxxxxxb"
DEV1="/dev/sdh1"
DEV2="/dev/sdh2"
MOUNT_POINT="/raid"
export PS1="[\T] [\W]# "
export JAVA_HOME=/usr/java/jre1.6.0_16
export EC2_HOME=/etc/ec2/apitools
export PATH=$PATH:$EC2_HOME/bin
export EC2_PRIVATE_KEY=/root/.ec2/pk-XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX.pem
export EC2_CERT=/root/.ec2/cert-XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX.pem
MAX_TRIES=10
# start/stop functions for OS
start() {
touch /var/lock/subsys/mountec2vol
INSTANCE=`curl http://169.254.169.254/latest/meta-data/instance-id 2> /dev/null`
CTR=0
/bin/echo "Mounting Elastic Block Store Volumes."
ec2-attach-volume $VOL1 -i $INSTANCE -d $DEV1
while [ ! -e "$DEV1" ]; do
/bin/sleep 1
CTR=`expr $CTR + 1`
if [ $CTR -eq $MAX_TRIES ]
then
/bin/echo "WARNING: Cannot attach volume $VOL1 to $DEV1 -- Giving up after $MAX_TRIES attempts"
exit 1
fi
done
ec2-attach-volume $VOL2 -i $INSTANCE -d $DEV2
while [ ! -e "$DEV2" ]; do
/bin/sleep 1
CTR=`expr $CTR + 1`
if [ $CTR -eq $MAX_TRIES ]
then
/bin/echo "WARNING: Cannot attach volume $VOL2 to $DEV2 -- Giving up after $MAX_TRIES attempts"
exit 1
fi
done
depmod -a
modprobe raid0
mdadm --assemble --verbose /dev/md0 /dev/sdh1 /dev/sdh2
if [ ! -d $MOUNT_POINT ]; then
mkdir $MOUNT_POINT
fi
/bin/mount /dev/md0 $MOUNT_POINT
}
stop() {
/bin/echo "Unmounting Elastic Block Store Volumes."
rm -f /var/lock/subsys/mountec2vol
/bin/umount $MOUNT_POINT
mdadm -S /dev/md0
ec2-detach-volume $VOL1
ec2-detach-volume $VOL2
}
case "$1" in
start)
start
;;
stop)
stop
;;
restart)
stop
sleep 5
start
;;
*)
echo "Usage: $0 {start|stop|restart}"
exit 1
esac
exit 0
This was for CentOS, but I imagine it would be easy enough to adapt to most other Linuxes. Change the Java version and remove the RAID attributes (and change the mount point) and you will be good to go. This would then be setup as an init script to run when the instance starts. An alternate way of accomplishing this would be to pass the EBS volume IDs as instance user-data, which would allow for a lot more flexibility.
It should be noted that this, of course, will not work for the root EBS volume - this setup presumes a mostly fixed root volume, with data kept on a separate mount-point. You can always set the EBS root volume from ec2-request-spot-instances using the --block-device-mapping parameter.
Best Answer
Your EC2 instance must be enabled for monitoring at 1-minute intervals. If you have not enabled detailed monitoring on your EC2 instance, then you would be collecting data in 5-minute intervals. 3 consecutive periods of 5-minute intervals would be 15 minutes.
I'm not certain, but after reviewing some of my own Cloudwatch alarms and playing with a new one in the console... it seems like, in this case, the alarm state triggers based on minutes instead of periods -- we just define minutes in terms of periods at alarm creation time. This seems sensible to me -- otherwise your alarm wouldn't ever be able to enter the alarm state if detailed (1-minute) monitoring was disabled.
Regarding detailed monitoring: I would turn it on for this case, if it is disabled. If you are using basic (5-minute) monitoring, the 3 data points don't necessarily mean that CPU Utilization has been >= 95% for 15 consecutive minutes. It rather means that CPU Utilization was >= 95% at the time the data was sampled, for 3 consecutive samplings.