A Framework to Design and Implement Real-time Multicore Schedulers using Machine Learning

Abstract

In this paper, we introduce a Framework to Design and Implement Real-time Multicore Schedulers using Machine Learning techniques applied to the very own data such systems produce as they operate. The framework builds on sensors and event counters present in modern hardware platforms and on variables kept by the operating system to capture run-time data that are subsequently subjected to ML tools to produce scheduling heuristics targeting specific optimization goals. It provides non-intrusive mechanisms to collect such data while the system runs real task sets with real workloads, thus preserving the quality of the captured data. It abstracts the Performance Monitoring Unit, thermal sensing, energy monitoring, and Dynamic Voltage and Frequency Scaling available on such platforms through a lean, architecture-independent API. After describing the framework in details, we demonstrate its applicability with the implementation of an energy-efficient, load balancing, real-time, multicore heuristic for a PEDF scheduler. The measured overhead imposed by the framework on the tasks it schedule is at most 0,0003583% and the maximum added jitter is less than 40μs, corroborating the ability of the framework to support the development of effective domain-specific schedulers using machine learning techniques.