MATLAB® Distributed Computing Server™ 3.3

Portable Batch System (PBS)

There are three versions of Portable Batch System (PBS) currently available:

• PBS Pro® or PBS Professional® is maintained and sold commercially by Altair Engineering.
• OpenPBS is the original open source version of the PBS scheduler and is available from www.openpbs.org  
• TORQUE is an open source extension to OpenPBS that is maintained by Cluster Resources.

Integration with Parallel Computing Toolbox™ and MATLAB® Distributed Computing Server™

Parallel Computing Toolbox and MATLAB Distributed Computing Server directly support PBS Pro and TORQUE schedulers. Using this scheduler, you can execute both distributed jobs, which do not have any intertask communication, and parallel jobs, which have intertask communication. Callbacks and dynamic addition of tasks to running jobs are not supported in all PBS variants.

The products provide example integration scripts that you can readily adapt for OpenPBS versions. All versions have the ability to manage multiple applications that include MathWorks parallel computing products. However, callbacks and dynamic addition of tasks to running jobs are not supported in PBS.

With the MathWorks job manager, the MATLAB Distributed Computing Server workers on the cluster run as a service and remain running at all times. With PBS, the workers run as an application that is started for the evaluation of tasks, and stopped when their tasks are complete. As a result, you incur an extra overhead of worker start-up and stop time. Create your tasks taking this overhead into account; for small tasks, this may represent a significant percentage of the overall execution time.

Features

PBS Pro offers several features, including:

• Scalability to very large heterogeneous clusters
• Support for resource usage policies, job priority and interdependency specification, and job accounting
• Automatic load leveling capability and enterprise-wide resource sharing and fault tolerance
• Support for several parallel programming libraries such as MPI, PVM and HPF
  

TORQUE offers several features, including:

• Scalability to very large clusters
• Support for resource usage policies
• Ability to manage serial and parallel jobs remotely
• Fault tolerance

Compared to PBS Pro and TORQUE, OpenPBS has a limited feature set.

TORQUE supports Linux® and UNIX® platforms only. PBS Professional also supports Windows® XP and Windows 2000 platforms .

All three have the ability to manage multiple applications that include MathWorks parallel computing products.

Information Resources

The following resources provide information on PBS versions:

• Documentation illustrates how you can use PBS to submit jobs for execution on your cluster in Programming Distributed Jobs and Programming Parallel Jobs.
• Vendor Web sites provide more information on schedulers:
  • PBS Pro or PBS Professional, available from Altair Engineering  
  • OpenPBS, available from www.openpbs.org  
  • TORQUE, available from Cluster Resources