HIPERCOMBATS : Overview of the parallel manager

The HIPERCOMBATS product has been designed

• to run efficiently on a non dedicated network of computers.
• to be used by mechanical engineers who have no experience in parallel computing,

In order to have a reliable and efficient parallel code, the parallel manager implement a dynamic load balanacing startegy and fault tolerances features. This allow to recover processes and computers breakdowns. Furthermore, the parallel aspect have been hidden as much as possible; the end-user does only access to the parallel capabilities of the code through the window where she/he selects the target computer.

This management window is also accessible at run time, when the parallel simulation is ongoing. It allows the user to dynamically change the computers she/he wants to use by deselecting or selecting new computers. This feature permits a large flexibility in the use of the HIPERCOMBATS product on a non dedicated network of computers. For instance for very long simulations, some workstations can be selected during the night and release during the working hours.
At run time, the user may inquire about the status of the simulation.

Once the parallel paramtric study is complete, a summary window is automatically open, where information on the simulation are displayed.
Below are the typical observed speedups on a network of heterogeneous SGI workstations (including 6 MPIS 5000 and a 4 processors MIPS 4400 Onyx). This basically show that the code is perfectly scallable (the sligthly worse speedup observed on more than 6 workstations are due to the addition of the less powerfull MIPS 4400 processors).

A Xpvm animation (1.7 MB) for a small parametric study can be displayed.