Performance analysis of the Kahan-enhanced scalar product onrncurrent multi-core and many-core processors

摘要

We investigate the performance characteristics of a numerically enhanced scalar product (dot) kernel looprnthat uses the Kahan algorithm to compensate for numerical errors, and describe efficient single instructionrnmultiple data-vectorized implementations on recent multi-core and many-core processors. Using low-levelrninstruction analysis and the execution-cache-memory performance model, we pinpoint the relevant performancernbottlenecks for single-core and thread-parallel execution and predict performance and saturationrnbehavior. We show that the Kahan-enhanced scalar product comes at almost no additional cost comparedrnwith the naive (non-Kahan) scalar product if appropriate low-level optimizations, notably single instructionrnmultiple data vectorization and unrolling, are applied. The execution-cache-memory model is extendedrnappropriately to accommodate not only modern Intel multicore chips but also the Intel Xeon Phi ‘KnightsrnCorner’ coprocessor and an IBM POWER8 CPU. This allows us to discuss the impact of processor featuresrnon the performance across four modern architectures that are relevant for high performance computing.