A Runtime Mapping Algorithm to Tolerate Permanent Faults in a CGRA

This work proposes to combine a fast runtime mapping algorithm to a fault tolerance mechanism to tolerate permanent faults in a Coarse-Grained Reconfigurable Architecture. In order to provide fault tolerance, the runtime mapping algorithm isolates the faulty functional units and allocates the instructions only in the fault-free ones. Since the solution depends on the amount of functional units available and the fault rate, fault injection and mapping were simulated considering different fault rates and architecture size. The results demonstrate that even at fault rates over 50% in functional units, the runtime mapping algorithm was able to map instructions into the architecture in most of the tested applications. Additionally, the time for mapping instructions into the CGRA remained in order of microseconds