Evaluation of Low-Cost Detection and Recovery of Soft Errors in an ABS controller
Paper in proceeding, 2009
This paper presents experimental results on a prototype brake controller's susceptibility to soft errors. We have developed a set of simple software-implemented error detection and recovery mechanisms which, in combination with hardware exceptions, aim at preventing soft errors from causing critical braking failures. To evaluate the effectiveness of these mechanisms, we injected roughly 60 000 single bit-flips into CPU registers and data memory of a MPC565 microcontroller running the brake controller program. The results show that the combined mechanisms effectively prevented critical braking failures; only three errors caused a critical behavior of the controller. As much as 69.0% of the errors escaped detection, but these errors had only a minor impact on the brake performance. Only 26.6% of the injected errors were detected and subsequently recovered from by the software. Of these were 91.5% detected by hardware exceptions and 8.5% by the software mechanisms. Despite all errors were injected into registers holding "live" data, as much as 46.9% were masked by the program and did not in any way affect the produced brake commands.