Stochastic thermal-aware real-time task scheduling with considerations of soft errors

• Feature the consideration of the uncertainty in transient fault occurrences.
• Enhance energy efficiency by selecting frequency via an energy efficiency factor.
• Task sequencing and DVFS are adopted to guarantee the system thermal constraint.
• Synthetic and real-life tasks are both utilized in two sets of simulations.

With the continued scaling of the CMOS devices, the exponential increase in power density has strikingly elevated the temperature of on-chip systems. Dynamic voltage/frequency scaling is a widely utilized system level power management technique to reduce the energy consumption and lower the on-chip temperature. However, scaling the voltage or frequency for thermal management leads to an increase in soft error rates, thus has adverse impact on system reliability. In this paper, the authors propose a stochastic thermal-aware task scheduling algorithm that considers soft errors in real-time embedded systems. For the given customer-defined soft error related target reliability and the maximum peak temperature, the proposed scheduling algorithm generates an energy-efficient task schedule by selecting the energy efficient operating frequency for each task and alternating the execution of hot tasks and cool tasks at the scaled operating frequency. The proposed stochastic scheduling algorithm features the consideration of uncertainty in transient fault occurrences. To handle the uncertainty, a fault adaptation variable α is introduced to adapt task execution to the stochastic property of fault occurrences. An energy efficiency factor δ is also introduced to facilitate the enhancement of energy efficiency by maximizing the energy saved per unit slack. Extensive simulations of synthetic real-time tasks and real-life benchmarking tasks were performed to validate the effectiveness of the proposed algorithm. Experimental results show that the proposed algorithm consumes up to 17.8% less energy as compared to the benchmarking schemes, and the peak temperature of the proposed algorithm is always below the maximum temperature limit and can be up to 9.6 °C lower than that of the benchmarking schemes.