Minotaur: Adapting Software Testing Techniques for Hardware Errors

摘要: With the end of conventional CMOS scaling, efficient resiliency solutions are needed to address increased likelihood hardware errors. Silent data corruptions (SDCs) especially harmful because they can create unacceptable output without user's knowledge. Several analysis techniques have been proposed identify SDC-causing instructions, but remain too slow for practical use and/or sacrifice accuracy improve speed. We develop Minotaur, a novel toolkit speed and analysis. The key insight behind Minotaur is that modern has many conceptual similarities software testing; therefore, adapting from rich testing literature lead principled significant improvements in identifies adapts four concepts testing: 1) it introduces concept input quality criteria PC coverage as simple effective criterion; 2) creates (fast) minimized inputs (slow) standard benchmark inputs, using assess goodness created input; 3) test case prioritization prioritize error injections invoke early termination given instruction up error-injection campaigns; 4) further or accelerate SDC discovery across multiple inputs. evaluate by applying Approxilyzer, state-of-the-art tool. Minotaur's first three Approxilyzer's 10.3X (on average) workloads studied. Moreover, 96% all instructions explored, compared 64% identified Approxilyzer alone. fourth technique (input prioritization) enables identifying explored at 2.3X faster than analyzing each independently our workloads.