Dynamic load balancing method on a heterogeneous cluster of workstations

The efficient usage of workstations clusters depends first of all on the distribution of the workload. The following paper introduces a method to obtain efficient load balancing for data parallel applications through dynamic data assignment and a simple priority mechanism, on a heterogeneous cluster of workstations, assuming no prior knowledge about the workload. This model improves the performance of load balancing methods in which one or more control processes remain idle for an extended period of time. In order to investigate the performance of this method we take into consideration a problem of 3D image reconstruction that arises from events detected by a data acquisition system. Studies of our load balancing model are performed under slight and heavy load condition. Experimental results demonstrate that this model yields a substantial load balance, even more if workstations are heavily loaded, from exploiting the idle time of one control process. In addition, this strategy reduces the overhead due to communication so that it could be successfully employed in other dynamic balancing approaches