MATLAB® Distributed Computing Server™ 3.3

Network Requirements

• All schedulers
• Local scheduler (in Parallel Computing Toolbox)
• MathWorks Job Manager
• mpiexec (shipped with MATLAB® Distributed Computing Server™)
• Platform™ LSF
• TORQUE
• Other Schedulers

Network Requirements for All Schedulers

For the interactive features - MATLAB® pool and Parallel Command Window - to work correctly additional TCP connections are opened by participating workers back to the client machine. This type of connection will not be possible if the client is firewalled or if the workers are on a private network with no external access. It is possible to control the client hostname and port used by the workers on both Windows and UNIX.

Network Requirements for the MathWorks Job Manager

• At least 5 gigabytes of disk space is recommended to accommodate the job manager's checkpoint directories
• Distributed computing processes must be able to identify each other by hostname, which requires the presence of a DNS or equivalent service on the network. In addition, the hostname with which a computer is visible to other computers must be the same as the hostname with which the computer identifies itself. For example, if a computer identifies itself as "computer1" then other computers must be able to resolve the name "computer1."
• The hostname for every node on the cluster must be bound to an IP address that matches an address on one of the host's network interface cards (NICs). Distributed computing processes will work correctly on machines with multiple NICs. See instructions for determining if your configuration is correct.
• Distributed computing processes make use of several TCP ports. On a node that runs a total of n job managers and workers, the MATLAB® Distributed Computing Server™ service reserves a total of 5 + n consecutive ports for its own use.
• Additional TCP ports are opened for inter-worker communication by participating workers during the execution of a parallel job. It is possible to control ports used for this purpose on both Windows and UNIX.
• If you use the job manager on a Windows cluster, transfer of large data sets may fail because of insufficient number of ephemeral TCP ports. This applies particularly to clusters with 32 or more workers. You must increase the default maximum number of ephemeral TCP ports set by Windows (5,000) on the computer running the job manager. This requires editing the Windows registry. For detailed instructions, see Troubleshooting in the MATLAB® Distributed Computing Server™ documentation.
• To run parallel applications that use inter-worker communication, all the cluster nodes running the application must have matching word sizes and processor endianness. As a result, heterogeneous cluster configurations running parallel applications can comprise only the following operating systems:
  • 32-bit Linux and Intel Macintosh
  
  Note that distributed applications that do not require inter-worker communication are supported for all combinations of operating systems in a computer cluster.

Network Requirements for mpiexec (shipped with MATLAB® Distributed Computing Server™)

• To run parallel applications that use inter-worker communication, all the cluster nodes running the application must have matching word sizes and processor endianness. As a result, heterogeneous cluster configurations running parallel applications can comprise only the following operating systems:
  • 32-bit Linux and Intel Macintosh

Network Requirements for Platform™ LSF

Users of Platform™ LSF 6.0 should apply Platform Computing's September 2005 maintenance pack or a subsequent maintenance pack.

Network requirements for TORQUE

When running parallel jobs under TORQUE, a remote execution program (such as SSH or RSH) must be configured on the cluster nodes, and submitting users must be able to use it without being prompted for any credentials.

Network Requirements for Other Schedulers

There are no specific requirements. Consult your scheduler's documentation for specific requirements.

MPI Requirements

Parallel Computing Toolbox™ ships with version 1.0.3 of MPICH2. The toolbox supports any MPI with an MPICH2-compatible "mpi.h" that is built as a shared library.