Comments



next up previous
Next: Conclusion Up: CS 267 Final Project: Previous: Performance

Comments

This project addresses a very narrow scope on the issue of parallelizing the solver of FEAP. For one, iterative solvers are easier to parallelize than direct solvers (though if you are using a parallel machine you are probably doing a large problem, and iterative solvers are ideal for large problems) - any FEM program should have at least one direct and one iterative solver.

As mentioned earlier the element state determination is not implemented in parallel. The element state determination - and the assembly of the stiffness, mass, and residual matrices - is not very complex to implement, though there are some interesting issues involved in the design of parallel algorithm.

There are some bugs in this implementation. The FEAP memory allocation system (a dictionary into a blank common), seems to not work perfectly on the slave nodes. This could be caused by the slave node's dictionaries not being initialized properly. The program also gets killed by the operating systems occasionally; this seems to happen consistently for a particular problem on a particular set of processors. For example, this test problem will die in the four processor case if nodes 4-7 are used, or the eight processor case which must use nodes 0-7 (though other problems were able to run on eight nodes), thus eight node timings are not reported,



ADAMS
Thu May 18 11:22:16 PDT 1995