HPAC: evaluating approximate computing techniques on HPC OpenMP applications

As we approach the limits of Moore's law, researchers are exploring new paradigms for future high-performance computing (HPC) systems. Approximate computing has gained traction by promising to deliver substantial computing power. However, due to the stringent accuracy requirements of HPC scientific applications, the broad adoption of approximate computing methods in HPC requires an in-depth understanding of the application's amenability to approximations. We develop HPAC, a framework with compiler and runtime support for code annotation and transformation, and accuracy vs. performance trade-off analysis of OpenMP HPC applications. We use HPAC to perform an in-depth analysis of the effectiveness of approximate computing techniques when applied to HPC applications. The results reveal possible performance gains of approximation and its interplay with parallel execution. For instance, in the LULESH proxy application approximation provides substantial performance gains due to the reduction of memory accesses. However, in the leukocyte benchmark approximation induces load imbalance in the parallel execution and thus limiting the performance gains.