Your scenario is correct in that terminating and restarting a new instance
charges you again, while stopping and restarting will not. Amazon charges you
at instance request time the entire hourly rate for that instance, whether you
start it at 12:01 or 12:59 and only get one minute of usage. Restarting just
uses the same instance that you already paid for, but you'll be charged again
at the top of the hour.
Because of this, Cloudman's autoscaling follows two basic rules:
1) Start instances up to the threshold set (max instances) if jobs are waiting
in the queue. We use the current job throughput rate and the number of queued
jobs to decide when to start an instance. For instance, if you have 1 job in
the queue and you're flying through jobs at a high rate, don't start a new
instance as it'd be a waste.
2) Kill all idle instances past the minimum threshold at 55 minutes past the
The reason for #2 is that if you've already paid for an instance, there's no
point in shutting it down early. You can always start it back up in less than
a minute as demand requires. We never stop a node. Once you've paid for it,
there simply isn't a reason to do this.
EBS volumes are a separate issue. Your cluster has several volumes attached to
the master node and shared with the workers via NFS. It does not matter how
many workers you have (other than data I/O costs, which are incredibly minor).
Let me know if I can provide more information. Thanks!
On Jan 24, 2012, at 4:05 PM, Jesse Erdmann wrote:
> We're in the planning phases of implementing an Amazon cluster for a
> particular workflow. We think volume will be such that it will make
> sense to have at least a the manager and a worker node running working
> all of the time and will be albe to reduce costs by using the reserved
> instance pricing. However, we have some questions about how auto
> scaling works with regards to stopping vs. terminating worker nodes.
> In my brief experience working with CloudMan I've been able to
> determine that allocating instances that use EBS volumes accrues
> charges. With the 15GB EBS volume currently allocated by the CloudMan
> AMI, starting a new worker node costs $1.50. So if I start a worker
> node, terminate it and start another one I get charged $3.00.
> However, if I start an instance, stop it and restart it I only get
> charged $1.50 for the initial allocation and I do not get charged for
> restarting it. Also, to my understanding (help me if I'm wrong here
> because my experience is limited) I don't get charged for the time
> where the volume is still mounted on EBS, but the instance is not
> running. I have a few questions regarding this.
> 1) For those with more experience with Amazon's pricing, is my scenario
> 2) Will the autoscaling of nodes terminate unused nodes or simply stop them?
> 3) If autoscaling terminates nodes, is it possible to stop and restart
> nodes within the cluster and not terminate them and still have
> CloudMan utilize them in the cluster appropriately?
> Jesse Erdmann
> Bioinformatics Analyst
> Masonic Cancer Center
> University of Minnesota
> Please keep all replies on the list by using "reply all"
> in your mail client. To manage your subscriptions to this
> and other Galaxy lists, please use the interface at:
Please keep all replies on the list by using "reply all"
in your mail client. To manage your subscriptions to this
and other Galaxy lists, please use the interface at: