摘要

This paper describes a technique for obtaining sums of floating point values that are independent of the order-of-operations, and thus attractive for use in global sums in massively parallel computations. The basic idea described here is to convert the floating point values into a representation using a set of long integers, with enough carry-bits to allow these integers to be summed across processors without need of carries at intermediate stages, before conversion of the final sum back to a real number. This approach is being used successfully in an earth system model, in which reproducibility of results is essential. Published by Elsevier B.V.

  • 出版日期2014-5