StarCluster 0.90

configuring, and managing a traditional computing cluster used in research labs or for general distributed computing applications.

StarCluster is built on top of EC2 which enables dynamically creating and destroying clusters of virtual machines and only paying for the time used. The amount per hour varies depending on the instance type and the number of virtual machines.

StarCluster consists of a library and set of scripts that use the library. For end-users, the scripts are the main user interface and provide simple intuitive options for getting started with distributed computing on EC2 (i.e. starting/stopping clusters, managing software configurations, etc). For developers, the library wraps the EC2 API to provide a simplified interface for launching/terminating nodes, executing commands on the nodes, copying files to/from the nodes, etc.

To get started, the user creates a simple configuration file with their account details and a few preferences (i.e. number of machines, instance type, EBS volumes to be mounted, etc). After creating the configuration file and starting the software, a cluster of Linux machines configured with a queuing system, a nfs shared /home directory, and OpenMPI is created and ready to go out of the box.

StarCluster has been targeted for computational research labs and to support classrooms with computational requirements.

For research labs, StarCluster is a way for graduate students and faculty to have an on-demand cluster. This means students can access their research with the same hardware and software configurations wherever they go; even if they move to another institution. StarCluster also provides a way for students to experiment with a computational model on a cheap budget before running on local dedicated resources.

In the classroom, StarCluster provides a cost effective, reliable way of managing the software configurations for a particular course. It also removes the majority of system administration concerns since the initial setup procedures have been captured in StarCluster and in the user’s software configurations (i.e. AMI images, EBS volumes, etc). This means that each semester the exact computing cluster configuration can be recalled with more or less nodes. With this model there is also the benefit that if hardware problems occur it’s easy to request a new set of machines in the cloud.