The Optimized Design of Stack for GRAPES's Adjoint Mode

The four dimensional variational data assimilation system (4DVAR) of GRAPES (Global/Region Assimilation and Prediction System) can use different meteorological data from different areas of different times obtained to optimize the quality of forecast based on an initialization background. As the core of the 4DVAR, tangent mode and adjoint mode can adjust the initialization background through using the deviation of the estimate of 3DVAR and observation.When a segment of the adjoint mode is run, the initial state of corresponding nonlinear mode might be needed as input. In order to balance the disadvantage of whole storage and whole computation, a double chained stack is used to store an interim data's snap for implementing the adjoint mode. Adopting the whole storage can speed up the adjoint mode prominently, but this may lead to the relation of first in and first out (FIFO) among some data blocks, which conflicts with the configuration of the double chained stack. A nested and double chained stack is proposed based on original double chained stack, using a kid chained stack to separate the data blocks that have FIFO relations. Data block pops first must be pushes in kid chained stack, and then can be popped at any time as needed. The nested and double chained stack can meet these requirements of different data blocks, FIFO or FILO, and satisfy the requirement of adjoint mode better. The result of experiment shows these approaches can double the operational speed with 30% extra memory.