Management Application - a New Approach to Control Many-Core Systems

Abstract

The increasing core count in many-core systems introduced management challenges, including scalability, portability, and reduced overhead in the user's applications. Works available in the literature seek to manage a given objective, such as power, temperature, communication, and quality-of-service. These management strategies are tightly coupled to the hardware platform and the operating system (OS) running on it. This coupling implies the lack of management modularity, resulting in low flexibility related to modifying management strategies at runtime, and low portability. State-of-the-art shows that few works propose management strategies or frameworks, only evaluating the proposed objective's quality. This work aims to present a new approach to control many-core systems, named Management Application (MA), which can implement multiobjective management decoupled from the hardware and the OS through a set of high-priority tasks. MA transforms the management problem into a distributed application, allowing the management to truly benefit from the high parallel power of many-cores. The MA approach is demonstrated with a proof-of-concept framework. Results evaluate the cost to adopt MA, compared to the cluster management, and the benefits of adopting MA to manage a benchmark with real-time constraints revealing improved memory footprint and higher management throughput due to its parallelization.