Dynamic load balancing on a network of workstations for solving computational fluid dynamics problems (Q1913188)

A methodology is presented for dynamic balancing of the computation load on a network of multi-user computers for parallel computing applications. In order to distribute the computation load in a multi-user environment, it becomes necessary to determine the effective speed of a multi-user workstation to a parallel application. In the present approach, it was assumed that (i) multi-user and multi-tasking networked computers may have different computation speeds, (ii) application data can be divided into many small data blocks with possibly different sizes, (iii) a process is assigned to each block, and (iv) the number of computers is much less than the number of processes. The developed dynamic load balancing procedure uses the greedy method for optimizing computation load distribution. Due to dynamic changes of the computer loads in a multi-user and multi-tasking environment, the loads on computers are periodically examined and parallel application processes may be re-distributed to reduce the computation time. The developed method has been tested on two computer clusters, and its applicability has been demonstrated for two case studies.