MARS: Middleware for Adaptive Reflective Computer Systems (2107.11417v1)

Abstract: Self-adaptive approaches for runtime resource management of manycore computing platforms often require a runtime model of the system that represents the software organization or the architecture of the target platform. The increasing heterogeneity in a platform's resource types and the interactions between resources pose challenges for coordinated model-based decision making in the face of dynamic workloads. Self-awareness properties address these challenges for emerging heterogeneous manycore processing (HMP) platforms through reflective resource managers. However, with HMP computing platform architectures evolving rapidly, porting the self-aware decision logic across different hardware platforms is challenging, requiring resource managers to update their models and platform-specific interfaces. We propose MARS (Middleware for Adaptive and Reflective Systems), a cross-layer and multi-platform framework that allows users to easily create resource managers by composing system models and resource management policies in a flexible and coordinated manner. MARS consists of a generic user-level sensing/actuation interface that allows for portable policy design, and a reflective system model used to coordinate multiple policies. We demonstrate MARS' interaction across multiple layers of the system stack through a dynamic voltage and frequency scaling (DVFS) policy example which can run on any Linux-based HMP computing platform.