Kinship: Resource Management for Performance and Functionally Asymmetric Platforms

On-chip heterogeneity has become key to balancing performance and power constraints, resulting in disparate (functionally overlapping but not equivalent) cores on a single die. Requiring developers to deal with such heterogeneity can impede adoption through increased programming effort and result in cross-platform incompatibility. To evolve systems software toward dynamically accommodating heterogeneity, this paper develops and evaluates the kinship approach and metric for mapping workloads to heterogeneous cores. For this metric, we provide a model and online methods for maximizing utility in terms of performance, power, or latency, to automatically choose the task-to-resource mappings best able to use the different features of heterogeneous cores. Such online scheduling at bounded cost is realized with a hypervisor-level implementation that is evaluated on a variety of actual, experimental heterogeneous platforms. These evaluations demonstrate the both general applicability and the utility of kinship-based scheduling, accommodating dynamic workloads with available resources and scaling both with the number of processes and with different types/configurations of compute resources. Performance improvements with kinship-based scheduling are obtained for runs across multiple generations of heterogeneous platforms.