Abstract:

As CMOS technology scales into the nanometer era, more transistors can be integrated in a single chip, which makes microprocessors become more vulnerable to highenergy particle induced soft errors. Quantitatively analysis of the Architectural Vulnerability Factor (AVF) of different structures is an important method to guide the microprocessor reliable design. Online AVF computing techniques utilize the features that AVF varies significantly across different workloads and individual structures, and then make a tradeoff between system reliability and performance. The paper proposes an occupancybased online AVF computing method, and evaluates it with a cycleaccurate simulator SimAlpha. Experimental results on SPEC CPU2000 integer benchmark suite show that the proposed method can compute AVF efficiently for different structures. Compared with an offline method, the averaged absolute errors of AVF for issue queue, reorder buffer, and load/store queue are 0.10, 0.01, and 0.039 respectively.

Key words: microprocessor;reliability;soft error;fault tolerance;architecture vulnerability factor