A Neural Network-Based Approach to Dynamic Core Morphing for AMPs

Asymmetric multicore processors (AMPs) are equipped with heterogeneous cores, each tailored with specific performance and energy characteristics. These cores are adept at handling the dynamic changes in computational demands that applications experience throughout their runtime, such as varying levels of instruction-level parallelism (ILP). Unlike their homogeneous counterparts, which are bound by a constant set of resources, AMPs can flexibly allocate tasks to the most energy-optimized core. ARM's big. LITTLE architecture is a notable example of this trend, combining out-of-order (OOO) and in-order (InO) cores on a single chip. However, while migrating threads dynamically between these cores can save energy, the overhead associated with such migrations can be a bottleneck. An emerging solution is the concept of morphable cores, which can switch between OOO and InO operational modes as needed. Despite minor overhead, this fine-grain adaptability is a signifi-cant step toward energy efficiency. However, the key challenge is to understand the application demands in real-time and select the optimal core accordingly. This work presents a neural network-based approach for core morphing that operates in real-time. Our results show that our approach achieves performance close to the roofline model for the targeted benchmarks.