Abstract
The increasing on-chip temperature has become a critical issue in the multiprocessor system on chip (MPSoC). It does not only disturb the reliability and performance of the system but it also increases the power consumption. Load balancing is a common scheduling technique to address the thermal issues where the load is transferred to relatively less used cores. However, during the transfer of workload, the destination cores may be in sleep mode and thus take some time to switch to the running mode. It results in wastage of some processing time. This paper presents a thermal-aware load balancing technique that avoids thermal emergencies while eliminating the switching delays. The proposed technique estimates the time taken by a task set to reach the temperature threshold values with the help of offline recorded thermal profiles of datasets. Moreover, cores are selected in a round-robin manner to equilibrate thermal gradients. The technique is based on Global Earliest Deadline First (GEDF) scheduling algorithm and considers ambient temperature, thermal cycles and energy consumption. The proposed technique is tested in a simulation environment comprising of scheduling and thermal simulation model. The results show that the technique reduces the average temperature up to 17%, thermal cycles up to 15%, energy consumption up to 20% and lowers the temporal and spatial gradients as compared to the commonly used predictive thermal-aware and thermal balancing techniques.