Pair and swap: An approach to graceful degradation for dependable chip multiprocessors

In this paper, we propose a processor-level fault tolerance technique called “Pair and Swap (P&S)” for a multi-core chip. In the P&S system, a 2n-cores-CMP (Chip Multiprocessor) which contains 2n processor cores composes n pairs. Two identical copies of a given task are executed on each pair of two processor cores and the results are compared repeatedly. If a fault is detected by a mismatch, partners of the mismatched pair are swapped with another pair and the mismatched task is re-executed from the latest checkpoint. Then, it is decided whether the fault is transient or permanent. If it is permanent, the faulty core is identified and isolated to reconfigure the entire system. P&S enables graceful degradation and tolerates both permanent and transient faults. We evaluate the performance of the proposed P&S and traditional triple module redundancy (TMR) using the Markov chains. The mean computation to failure of the P&S is about 1.4 times larger than that of dynamic TMR scheme.