Abstract
In this paper we study a new strategy for adaptive dynamic load balancing in large heterogeneous distributed systems. This strategy that aims to minimise the application response times comprises two phases, each of which is simple enough to be implemented as a μKernel-level service. In the first phase, unlike in most dynamic load balancing algorithms developed so far, we take into account the different types and sizes of parallel applications that are run on a general purpose parallel computer in adaptively deciding an execution domain for each new application. In the second phase, a classical Receiver Initiated Diffusion technique is used to balance the load among neighbour nodes in the system. Using a very simple, yet efficient classification of the parallel applications into computation hard, communication hard and balanced, we simulated the behaviour of a simplified distributed system, obtaining results that justify the employment of such a two phase allocation algorithm as a μKernel service for large heterogeneous distributed systems.
| Original language | English |
|---|---|
| Place of Publication | Amsterdam, Netherlands |
| Publisher | IOS Press |
| Number of pages | 11 |
| Publication status | Published - 1 Dec 1995 |