How to improve local load balancing policies by distorting load information

Abstract

The paper focuses on local load balancing policies for massively parallel architectures and introduces a new scheme for load information exchange between neighbor nodes. The idea is to distort the exchanged load information to let the policy keep into account a more global view of the system and overcome the limits of the local scope. The presented scheme has been integrated into two variants of a direct-neighbor policy and evaluated in dependence of the characteristics of the system load. Experimental results show that the transmission of distorted load information provides high efficiency unless the dynamicity of the load becomes too high, in which case it is preferable to exploit non-distorted load information.