StarCluster - Mailing List Archive

Re: Smart Cluster

From: MacMullan, Hugh <no email>
Date: Tue, 29 Jan 2013 14:31:19 +0000

Hi Sergio:

a -- yes!

MASTER_INSTANCE_TYPE = c1.medium
MASTER_IMAGE_ID = ami-XXXXXX
NODE_IMAGE_ID = ami-XXXXXX
NODE_INSTANCE_TYPE = cr1.8xlarge

b -- in config do:

In the [cluster NAME] section:
DISABLE_QUEUE=True
PLUGINS = sge

Then define your sge plugin with:

[plugin sge]
setup_class = starcluster.plugins.sge.SGEPlugin
master_is_exec_host = False

You can also force the master to be a spot instance by using 'starcluster start --force-spot-master', which can save serious cash (90%).

-Hugh


On Jan 29, 2013, at 9:02 AM, "Sergio Mafra" <sergiohmafra_at_gmail.com<mailto:sergiohmafra_at_gmail.com>> wrote:

Hi fellows,

If we are supposed to have the instances running for a long time, LoadBlancer can help us a lot, turning on and off when is needed.
I wonder if we can make this more smarter and cheaper, launching the master node as a m1.small instance, taking care only for communication and queue managemente, leaving the heavier job for the compute nodes, that can take advantage of cluster type instance and placement groups.
So my question is:

a) Can we have a cluster with differente instance types - Master and Compute nodes?
b) How to exclude Master node from being a computacional node in the job queue?


All the best,

SÚrgio
_______________________________________________
StarCluster mailing list
StarCluster_at_mit.edu<mailto:StarCluster_at_mit.edu>
http://mailman.mit.edu/mailman/listinfo/starcluster
Received on Tue Jan 29 2013 - 09:31:25 EST
This archive was generated by hypermail 2.3.0.

Search:

Sort all by:

Date

Month

Thread

Author

Subject