ML for System-Level Modeling

Ever-increasing complexity and heterogeneity of systems and applications pose fundamental new challenges for design, programming, and runtime management of compute systems. Fast and accurate models of architectures and the applications running on them are essential for early design space exploration and hardware/software co-development. Traditional simulation-based or analytical models are often too slow or inaccurate to effectively support design processes. Machine learning for system-level modeling has emerged to bridge the gap between slow simulation-based methodologies and inaccurate analytical models. The work in this space is large and diverse, making predictions for different aspects of the system. We define a three-dimensional space to differentiate existing work in system-level models based on their inputs and target predictions: (1) Cross-layer models predict low-level architectural details based on high-level features, allowing architects to estimate power and performance of real-world applications that are too complex to run on a detailed simulator; (2) cross-platform models learn the relationship of program executions on different platforms, allowing programmers to profile their code on a development machine and estimate the expected behavior on other machines; and (3) runtime workload models forecast future workload behaviors by looking at their execution history, enabling proactive runtime resource management. In this chapter, we describe representative works in each dimension and discuss their trade-offs and accuracies.