HyperSizer White Paper

"Parallel Processing with HyperSizer" (9 pages)


Phil Yarrington, Craig Collier, Mark Pickenheim


Parallel Processing. A parallel capability has been developed to dramatically reduce the run-times of HyperSizer analysis and optimization of large finite element models. The parallel computing system uses computers running Windows XP/2000/NT and was designed to take advantage of both multiprocessor workstations and workstations networked together. The parallel capability has the following features: A central controlling (client) computer divides a large finite element model into pieces (called groups) and sends each group to a different processor or workstation (server) for processing. The server workstations, rather than the client, perform the majority of HyperSizer I/O and numerical processing so that for long running processes, a high parallel to serial computing ratio is achieved; Only one HyperSizer parallel server license is required to execute HyperSizer on as many computers and processors as desired; Network traffic is minimized between computers by allowing each workstation to perform its own disk I/O. This means the HyperSizer parallel process does not bog down the network; The HyperSizer processes run as "low-priority" tasks, so that the parallel processes have very little impact on workstation interactive users; For long running analyses and optimizations, the overhead of distributing processes is very nearly "lost in the wash" and the parallelism is very efficient yielding nearly one-to-one processor to speedup ratios