Cooperative user- and system-level scheduling of task-centric parallel programs

Abstract: Emerging architecture designs include tens of processing cores on a single chip die; it is believed that the number of cores will reach the hundreds in not so many years from now. However, most common workloads cannot expose fluctuating parallelism, insufficient to utilize such systems. The combination of these issues suggests that large-scale systems will be either multiprogrammed or have their unneeded resources powered off. To achieve these features, workloads must be able to provide a metric on their parallelism which the system can use to dynamically adapt per-application resource allotments.Adaptive resource management requires scheduling abstractions to be split into two cooperating layers. The system layer that is aware of the availability of resources and the application layer which can accurately and iteratively estimate the workload's true resource requirements.This thesis addresses these issues and provides a self-adapting work-stealing scheduling method that can achieve expected performance while conserving resources. This method is based on deterministic victim selection (DVS) that controls the concentration of the load among the worker threads. It allows to use the number of spawned but not yet processed tasks as a metric for the requirements. Because this metric measures work to be executed in the future instead of past behavior, DVS is versatile to handlevery irregular workloads.