Experiences in dynamic placement of actors on multicomputer systems

The availability of low-cost microprocessors and the advances in communication technologies have promoted a great interest in multicomputer architectures. These systems are characterized by high performance, availability and extensibility at low cost but nowadays lack of good programming tools which could help the programmer to really exploit the resources of the computing system. One of the most significant problems, especially for highly dynamic computations, is how to allocate work to the nodes of the system. We examine the main topics concerning a dynamic load balancer integrated in the ASK run-time kernel which supports parallel actor programming on a transputer network. The balancer implements both a random allocation algorithm and a more complex one called Adaptive Contracting Within Neighbourhood (ACWN). Experimental results concerning the balancer behavior on some sample algorithms are presented and analyzed