Setting up an HPC cluster in the cloud can be a daunting task for new users looking to utilize the cloud to run their HPC applications. Learning the ins and outs of the infrastructure as a service (IaaS) model in addition to configuring and installing a typical HPC system is not an easy task.
In order to use the cloud effectively users need to be able to automate the process of requesting and configuring new resources and also terminate resources when they’re no longer required without losing data. These concerns can be a challenge even for advanced users and require some level of cloud programming in order to get it right. In an effort to improve this situation, the Software Tools for Academics and Researchers (STAR) group at MIT has created an open-source project called StarCluster that allows anyone to create and manage their own HPC clusters hosted on Amazon’s Elastic Compute Cloud (EC2) without needing to be a cloud expert.
One of StarCluster’s primary goals is to be simple to use and to hide as many of the cloud computing details from users as possible. When a new user attempts to use StarCluster for the first time an example configuration file is created that is ready to be used out-of-the-box. The user simply needs to fill in the EC2 account information and optionally customize the number of machines to use before he or she is ready to start a cluster. Starting a cluster with the example configuration will launch a two-machine cluster using the cheapest instance types available on EC2. This allows users to experiment with StarCluster for the first time without dramatic up-front costs.
The group of cluster-specific settings in the configuration file is known as a “cluster template”. StarCluster supports defining multiple cluster templates which can be used when launching a cluster. For example, it’s often useful to have separate templates for different cluster sizes such as a template that defines a small two-machine cluster and another template that defines a large ten-machine cluster. These templates can be specified at runtime to allow a variety of configurations to be used when starting a cluster.
Starting an HPC Cluster on EC2
Once the configuration file has been created, starting a cluster is as simple as running “starcluster start mynewcluster” at the command line. This command will first verify that all settings in the configuration file are valid and are likely to create a working system. Once the settings in the configuration file have been verified, the “start” command creates a new cluster based on these settings with a tag-name of “mynewcluster” on EC2.
Once the “start” command has finished the user can login to the “master” machine as root by running “starcluster sshmaster mynewcluster”. At this point the user has the (root) keys to the cluster just as they would with their own local resources.
StarCluster also has the ability to create multiple HPC clusters. Running the same “start” command again with a different tag-name will launch another HPC cluster in the cloud using the same settings as the previous run. If you’ve defined additional cluster templates in the configuration file these can optionally be used to specify a different group of settings to use when starting the next cluster.
Once the user has finished using a cluster they simply specify its tag-name to StarCluster’s “stop” command to shut it down. For the “mynewcluster” example above the command would be “starcluster stop mynewcluster”. The “stop” command will shutdown the entire cluster and terminate the billing period.
Automated HPC Cluster Configuration
StarCluster automatically configures each machine with the appropriate networking settings needed to communicate with the rest of the cluster. On top of this, StarCluster also fully configures password-less SSH communication for both the root user and a normal user on the cluster. Password-less SSH allows a user to login remotely between machines in the cluster without using a password. This is useful when administering the machines in the cloud and is also a necessary requirement for OpenMPI communication.
Most clusters usually have some form of a queuing system for submitting and load-balancing many computationally intensive tasks or “jobs” and StarCluster is no exception. Out-of-the-box, StarCluster installs and configures the open-source version of the Sun Grid Engine (SGE) queuing system for running distributed and parallel jobs on the cluster. A parallel queue is also configured by default that enables SGE to monitor and account for parallel tasks that use more than one machine in a single job.
Many parallel tasks are commonly written using the Message Passing Interface (MPI). For MPI users, StarCluster includes an SGE-aware OpenMPI installation that provides tight integration between the SGE job scheduler and MPI applications. This integration removes the need for users to specify a list of hosts to use when running an MPI job. Rather, OpenMPI will automatically fetch the host info it needs directly from SGE and begin execution. This allows all machines involved in the MPI calculation to be correctly accounted for by the queuing system.
Sharing files between machines without manually copying files around is a requirement for most HPC systems. Typically this is done using a shared folder via the network file system (NFS). StarCluster automatically configures /home on each “worker” machine of the cluster to be NFS-shared from the “master” machine. This allows users to see their files on any machine in the cluster and also provides a globally accessible place for jobs to read input data and write their finished results.
The StarCluster Amazon Machine Image (AMI)
Amazon Machine Images are used by EC2 to load an entire operating system along with various applications, libraries, and data onto a newly requested virtual machine. Machine images are publicly available for just about any Linux distribution, Solaris, and even Microsoft Windows. New images can be created with custom software configurations by launching a new virtual machine from an existing AMI, installing your new software, and then running an AMI creation process on the machine to create a new AMI.
StarCluster comes with a publicly available custom-tailored AMI, in both 32bit and 64bit flavors, that contains the entire OS and software configuration needed for an HPC cluster on Amazon. The StarCluster AMI is Ubuntu Linux 9.10 based and includes the Sun Grid Engine queuing system (open-source edition), the network file system, and OpenMPI along with common development tools and libraries to compile new software from source. The StarCluster AMI also includes a custom-compiled installation of the Automatically Tuned Linear Algebra Subroutines (ATLAS) and Linear Algebra PACKage (LAPACK) libraries that have been optimized for the larger high-CPU instance types on EC2. For numerical python users, the AMI contains both NumPy and SciPy installations that have been custom compiled against the optimized LAPACK/ATLAS installations. These optimized libraries provide a significant performance improvement when running linear algebra routines in the cloud.
Of course, StarCluster does not limit you to only these software installations. The StarCluster AMIs can easily be extended with your own software to create a brand-new AMI tailored for a specific need. To simplify the AMI creation process StarCluster provides a “createimage” command that will automatically create a new AMI from a running Amazon EC2 virtual machine in the cloud. This allows you to launch a single virtual machine, install your software, and easily create a new AMI from this machine. Using a new customized AMI with StarCluster is as simple as updating the configuration file with the new AMI’s identifier.
Using EBS Volumes for Persistent Storage
Amazon also provides a service called Elastic Block Storage (EBS) which allows users to create virtual block storage volumes that are similar in functionality to a USB pen-drive. These volumes can be anywhere from 1GB to 1TB in size and can be attached to a single virtual machine in the cloud at a time. The benefit of using these volumes is that any data written to EBS is automatically stored and persisted in the cloud even after all virtual machines have been terminated. This means the next time you start a cluster and attach the EBS volume, all of your data will be available as it was the last time you launched a cluster. Another benefit of using EBS volumes is that they’re easy to snapshot and duplicate which allows for backing up large amounts of data in the cloud.
StarCluster has the ability to utilize Amazon’s EBS volumes to provide persistent data storage for a given cluster. To use EBS with StarCluster you must first create an EBS volume. For new users, this process is simplified by using StarCluster’s “createvolume” command. This command automates the process of creating, partitioning, and formatting a new EBS volume.
Using a new volume with StarCluster involves adding additional volume settings to the configuration file. These settings specify the volume to use and the location on the cluster’s file system to attach the volume. This file system location is then NFS-shared from the “master” machine to all “worker” machines. StarCluster does not limit you to using a single EBS volume. Multiple EBS volumes can be configured, attached, and shared on the cluster. This allows up to several terabytes of data to be stored on the cluster.Getting Started with StarCluster
StarCluster is open-source software and can be downloaded for free from the StarCluster website at http://web.mit.edu/starcluster or from the Python Package Index (PyPI) at http://pypi.python.org/pypi/StarCluster.
UPDATE: We now have a video screencast of StarCluster in action that can be viewed here.
About the Author
Justin Riley is a software developer for the Software Tools for Academics and Researchers (STAR) group at the Massachusetts Institute of Technology (MIT). The STAR group seeks to bridge the divide between scientific research and the classroom by collaborating with faculty from MIT and other educational institutions to design software that explores core scientific research concepts. The STAR group works out of the Office of Educational Innovation and Technology (OEIT) under the Dean for Undergraduate Education (DUE) at MIT.
Justin has been developing with the Amazon cloud for the past three years and has successfully used the cloud to support the “Introduction to Modeling and Simulation” and “Intro to Parallel Programming for Multicore Machines using OpenMP and OpenMPI” courses at MIT. His work with StarCluster came directly from the need to provide a sustainable solution to the issues associated with bringing computational resources into the classroom. Justin created StarCluster to automate the process of locating, configuring, and maintaining computational resources without needing to be a 24/7 system administrator and without having to make a physical appearance to address potential hardware and software issues.