Abstract

Many load balancing techniques utilize graph representations of tasks and their communication patters. The properties of these graphs and the characteristics of interest heavily impact the design and implementation of the underlying algorithms. To successfully evaluate the design space of these algorithms a generic, composable, and performant library is needed. In this talk I present the Task Scales library, the new Normalized Weight algorithm I have developed (along with it’s supporting data structures), and the abstract model that supports the relational partitioning necessary for topologically aware load balancing.