The large ensemble data assimilation computations were only possible using the leading-edge K computer and the eigenvalue solver "EigenExa" that allows an extremely effective use of the K computer. We achieved amazingly high 44% efficiency, 263 TFLOPS using 4608 nodes of the K computer (about 1/20 of the full capacity). The parallel-efficient LETKF was also a must.
With 10240 members, we could obtain a very precise probabilistic representation of the earth atmosphere. Long-range error correlations beyond continental scales and bimodal structures of moisture variables were clearly represented. These were very difficult to observe with less than a few hundred members, a typical choice in ensemble data assimilation of the global atmosphere.
For more details, refer to the press release on this research achievement. Here are the links to the press release and the original research article published in Geophysical Research Letters.
- Miyoshi, T., K. Kondo, and T. Imamura, 2014: The 10240-member ensemble Kalman filtering with an intermediate AGCM. Geophys. Res. Lett., 41, doi:10.1002/2014GL060863.