Linpack并行性能模型及其预测

摘要

通过分析大规模Linpack的主要时间开销,建立关于矩阵规模、网络带宽、下三角方程求解效率以及矩阵乘法效率的Linpack并行性能模型.在神威蓝光和神威4000A上分别对该模型进行可信性验证,结果证明,该模型能较好地估计主要因素的改变对Linpack并行性能的影响,尤其在大规模矩阵情况下预测误差不到1％.%Based on the most time consumption consists of decomposition, update and backward substitution, this paper proposes a Linpack parallel performance model about matrix scale, network bandwidth, triangle equation solution efficiency and matrix multiplication efficiency for quickly predicting the Linpack parallel performance. It validates this model on the Sunway Bluelight and Sunway 4000A. Results testify the creditability of the model, and show it can well estimate main factors to change the influence of Linpack parallel performance, especially at the massive matrix scale, the error is less than 1%.