Tech Briefs




Implicit Solutions with Distributed Memory Processing (DMP)

A distributed memory version of the ADINA implicit solver is now available for the sparse solver (Solids, CFD, and FSI) and the 3D-iterative solver (Solids and FSI). The DMP solver partitions the model matrix among the available CPUs. The processors work in parallel to split the computational effort and the total solver memory required for solution. For large problems, the reduced memory requirements per CPU can lead to much bigger models fitting in physical memory (in-core). In its current version, most of the benefits of the implicit DMP solver are realized for this class of problems, namely those which would otherwise not fit in core memory.

We present here one example to illustrate the benefits of using the DMP solver. The industrial component simulated is shown in the figure above. The model has slightly over 5 million degrees of freedom with mesh gluing, and is highly nonlinear due to the presence of over 100 bolts, multiple contact surfaces with friction, and large displacements. The model is subjected to a bolt-preload step, followed by applied external loading.

The model was solved with the 3D-iterative solver on a cluster of 'early generation' Intel Xeon x86-64 workstations connected via a low-cost commodity Gigabit Ethernet switch. The DMP run used 4 computers (each with dual processors), for a total of 8 CPUs.

The storage requirement for the model data and solver is easily accommodated within the total available memory of the 4 computers in the cluster. The total solution time was 6.4 hours. A comparable analysis with one of the workstations would have taken considerably more time to solve, mainly due to the out-of-core solution required. The von Mises stresses at the end of the simulation are shown in the figure below. Note how the high stresses are mostly localized to the bolted areas.






This new implicit DMP solver and the new explicit DMP solver, highlighted in a previous newsletter (ADINA News, Oct. 15, 2007), clearly already strengthen ADINA for the solution of very large computational models. However, we are continuing with our development of the DMP version of ADINA to obtain even greater effectiveness.