Research On Fault Recovery Techniques For Soft Errors Of COTS DSP

The Digital Signal Processor(DSP)has been widely used in the field of aerospace,especially in the on-board system.However,transient faults caused by high energy particle irradiation can seriously affect the reliability of DSP.Although the anti-radiation device can effectively resist the impact of the spatial particle irradiation,but at the same time,it will introduce huge cost in power consumption and performance.Compared with the hardware anti-irradiation devices,improving the reliability of Commercial Off-The-Shelf(COTS)DSP by using software fault tolerance technologies has obvious advantages in terms of performance,power consumption,cost and efficiency of development.So it is important to study the software fault tolerance technologies based on DSP COTS platform for promoting the development of China's aerospace industry.Based on the analysis of existing software fault tolerance technologies,In this paper,we study the software error recovery technology for the COTS DSP linear assembly program.The main work of this paper are summarized as follows:1)We propose an efficient approach to detect and correct control-flow errors for the linear assembly(DCCLA).Based on the characteristics of linear assembly,DCCLA proposes the concept of loop block and non-loop block.DCCLA could detect intra-block errors,inter-block errors and inter-function errors by using the mechanism of instruction counting and label distribution.DCCLA uses hierarchical processing to achieve fault recovery.In addition,DCCLA also proposes three optimization strategies for different performance and reliability constraints.2)We propose a fine-grained fault recovery technique based on idempotent(FFRBI).FFRBI divides program into idempotent regions based on the analysis of idempotency.Then FFRBI could detect data flow errors by using the mechanism of instruction redundancy.FFRBI could correct data errors by re-executing idempotent region with small performance overhead.In addition,FFRBI can be combined with other data error detection algorithms conveniently to achieve configurable fault tolerance.3)We expand the fault tolerance automatic conversion tool(FTACT)through the implementation of DCCLA and FFRBI.FTACT could automatically reinforce programs based on DCCLA or FFRBI.4)We design and implement the tool of random fault injection based on debug server scriping(RFIDSS)designed and implemented.The effectiveness of DCCLA and FFRBI methods is verified by high energy particle irradiation experiments and simulated fault injection experiments.In addition,theperformance overhead of the reinforcement is analyzed based on the CCStudio compiler.