High-Performance and Energy-Efficient Fault Tolerance Scheduling Algorithm Based on Improved TMR for Heterogeneous System
Shigan Yu, Zhimin Tang, Xiaochun Ye, Zhimin Zhang, Dongrui Fan, Zhiying Jiang · 2018
Traditional TMR (Three Mode Redundancy) method based on homogeneous multicore has been used in solving the transient fault of microprocessor and improving reliability, but it is characterized by low performance and high power consumption. This paper proposes an improved TMR method to improve the performance of fault-tolerant scheduling for heterogeneous system (HEFT-ITMR). The tasks are divided into two groups with/without fault-tolerant requirements, which are scheduled by traditional TMR method and by competitive mechanism, respectively. The tasks without fault-tolerant requirements will be re-executed in traditional TMR method if the task reliability fail to meet the requirements. The experimental results demonstrate that the proposed method can achieve a 18.1% improvement in average efficiency over traditional TMR method when running three types of benchmark without injecting errors and yield an average speedup of 15.3% and the power consumption can also be decreased by 20.5% over TMR when 200 errors are injected. In addition, a parallel optimization method P-HEFT-ITMR is proposed to further improve system performance and reduce system power consumption while ensuring system reliability.