SVM-Based Dynamic Voltage Prediction for Online Thermally Constrained Task Scheduling in 3-D Multicore Processors

Abstract

Hotspots occur frequently in 3-Dmulticore processors (3D-MCPs) and they may adversely impact the reliability of the system and its lifetime. We present support-vector-machine (SVM)-based dynamic voltage assignment (SVMDVA) strategy to select voltages among low-power and high-performance operating modes for reducing hotspots and optimizing performance in 3D-MCPs. The proposed SVMDVA can be employed in online, thermally constrained task schedulers. First, we revealed two different thermal regions of 3D-MCPs and extract different key features of these regions. Based on these key features, SVM models are constructed to predict the thermal behavior and the best operation mode of 3D-MCPs during runtime. SVMDVA using SVM models with monitoring workload and temperature behavior can effectively limit the temperature increase in 3D-MCPs. This is extremely important for accurately predicting the thermal behavior and providing the optimum operating condition of 3D-MCPs to achieve the best system performance. Experimental results show that SVMDVA is an effective technique for reducing hotspot occurrences (57.19%) and increasing throughput (25.41%) for 3D-MCPs.