SNNS for Workstation Clusters

Training several neural networks on a single workstation may require a large amount of training time, if performed on a single workstation. Dividing the training task among several workstations on a LAN (local area network) may significantly decrease the overall training time. This version of SNNS for workstation clusters allows to train several (possibly different) networks on different workstations and to control them with a single graphical user-interface.

The kernel interface has been extended for this purpose by an rpc module that is controlled by the graphical user-interface xgui. The module was named rpc because the communication among the workstations is performed via the rpc (remote procedure call) protocol.

This parallel SNNS version supports the following methods of parallelism:

• training different networks in parallel, each with its own pattern file,
• training several copies of the same network with different parameters or different initialization in parallel, each with its own (copied) pattern file, or
• an arbitrary combination of the above two methods.

The extension consists of two programs:

• xguirpc: SNNS simulator program (xgui, simulator kernel, and rpc module linked into one executable program). This program is the graphical user-interface, which controls the different simulator kernels. It has the same functionality as the xgui (snns) program plus different panels to control the rpc-interface. To use multiple simulator kernels, one may start as many kernelrpc programs as desired.
• kernelrpc: simulator kernel linked with the rpc module (the server). This program is the stand alone simulator kernel with rpc module, which is controlled from the xguirpc program. Each kernel has is own net, pattern sets and configurations.

The files are located in the directory

<snns-root>/rpc/bin/<architecture>

where <architecture> is one of the currently supported machine architectures. We have tested it on an Intel Paragon and single CPU Sun SparcStations, and Linux systems. To use the program efficiently on multi-CPU systems like the DEC Alpha some adjustments in the libraries might be necessary.

The easiest way to start a new simulator kernel is from the new panel called by pressing the button in the SNNS rpc panel (see ).

If the networks are to be trained without human supervision, a similar functionality may also be obtained by starting a copy of the batch SNNS simulator snnsbat on each workstation.

• SNNS RPC Programs
  • Xguirpc
  • Kernelrpc
• Control Panels
  • SNNS RPC Panel
  • RPC Setup Panel
  • New Panel
  • Transfer Panels
  • Show Info Panel
• RPC Errors