On soft errors in the Conjugate Gradient method: sensitivity and robust numerical detection -revised

Overview & Synopsis

The conjugate gradient (CG) method is the most widely used iterative scheme forthe solution of large sparse systems of linear equations when the matrix is symmetric positivedefinite. Although more than sixty year old, it is still a serious candidate for extreme-scalecomputation on large computing platforms. On the technological side, the continuous shrinkingof transistor geometry and the increasing complexity of these devices affect dramatically theirsensitivity to natural radiation, and thus diminish their reliability. One of the most common effectsproduced by natural radiation is the single event upset which consists in a bit-flip in a memory cellproducing unexpected results at application level. Consequently, the future computing facilitiesat extreme scale might be more prone to errors of any kind including bit-flip during calculation.These numerical and technological observations are the main motivations for this work, where wefirst investigate through extensive numerical experiments the sensitivity of CG to bit-flips in itsmain computationally intensive kernels, namely the matrix-vector product and the preconditionerapplication. We further propose numerical criteria to detect the occurrence of such soft errors; weassess their robustness through extensive numerical experiments.