Power- Thermal Aware Balanced Task-Resource Co-Allocation in Heterogeneous Many CPU-GPU Cores NoC in Dark Silicon Era

To provide balanced mapping for multiple applications in many-core heterogeneous CPU-GPU systems on network-on-chip (NoC) in dark silicon era, power-thermal aware task-resource co-allocation with reconfigurable NoC framework is proposed in this work. Task allocation and resource configuration problem is initially formulated using linear programming (LP) optimization model to search for an optimal solution. Distributed resource management in heterogeneous CPU-GPU systems with a fast balanced mapping heuristic is proposed to minimize power and thermal hotspots and improve performance at run-time while meeting the computation, communication, power and thermal budgets constraints of manycore NoC in dark silicon. We have also implemented a state-of-the-art minimum-path contiguous mapping for comparisons. Simulations under real-world benchmarks and platforms show that the proposed dynamic load-balanced mapping strategy improves NoC latency and throughput by 50-100% (while providing near-optimal solution) compared to minimum-path.