|
Abstract : |
Abstract: We consider the problem of mapping tasks to processor nodes at run-time in multiprogrammed multicomputer systems (i.e. message passing MIMD-systems). Besides load balancing, the goal is to place intensively communicating tasks close together to minimize communication delays. Since the placement has to be done at run-time by migrating tasks, migration cost are also considered. Our decentralized approach is inspired by a physical analogy: Tasks are considered as particles acted upon by forces. Each aspect of the allocation goal is modeled by a dedicated force. Migration activities cease, when a stable situation with balanced forces is reached. Simulations for multiprogrammed hypercube and 2D-grid topologies confirmed the usefulness and suggests superiority of this more general approach to other decentralized algorithms for the environment considered. 1., |
