Distributed control parallelism in multidisciplinary aircraft design

A fully‐distributed parallel global search algorithm

“…The distributed version with threads performs the worst of all the methods, rapidly decreasing in efficiency as the number of processors used is increased. This behaviour was not observed for pthreads on the Intel Paragon reported in Krasteva et al (1999), and thus is more likely a reflection of the SGI pthreads implementation than of an inherent characteristic of pthreads. A detailed discussion of these parallel performance results, relative to the SGI Origin hardware and process assignment to memory banks and CPUs, can be found in Baker (2000).…”

“…It was observed that the master-slave load balancing method was the most efficient for a large number of processors, when the variation in function evaluation times was small (because the master then did little redistribution work). When the variation in function evaluation times is significant, as is the case for the aggressive DIRECT algorithm or inherently in other aircraft design problems (Krasteva et al (1999)), or as here when using a small number of processors, the fully distributed dynamic load balancing method is most efficient. The implication is that for large scale realistic MDO problems, compared to static distribution or dynamic load balancing via a master-slave paradigm, the fully distributed control paradigm for dynamic load balancing will scale the best to massively parallel (distributed memory) machines.…”

“…The dynamic load balancing algorithm is based on that of previous work (Krasteva et al (1999)), employing random polling for the redistribution of tasks and token passing to terminate the load balancing process. Once task evaluation is started by a processor, it evaluates a single task and then processes any messages received during the evaluation of the task.…”

“…The termination detection scheme used for DLBDC and DLBDCT is the standard token wave algorithm used in Krasteva et al (1999). Suppose there are P processors.…”

“…The interprocessor communications used for the DIRECT box manipulation by the fully distributed version of dynamic load balancing are the same as those performed by the static version, with the added capability of task migration to processors that have finished their tasks. The dynamic load balancing algorithm is based on that of previous work (Krasteva et al, 1999), employing random polling for the redistribution of tasks and token passing to terminate the load balancing process. Once task evaluation is started by a processor, it evaluates a single task and then processes any messages received during the evaluation of the task.…”