University of California Los Angeles

We have developed a new parallel algorithm for EnKF in conjunction with AOML HRD. The algorithm relies upon two different methods that are complimentary to one another: covariance localization, where the ensemble covariances are restricted as a function of distance; and local analysis, where each analysis point is computed using only the variables within range. We analyze the mathematical properties of these two different approaches and determine when and where each approach is most beneficial. We also look at the pros and cons of solving, in parallel, for all observations or all model points first, and conclude that when the length-scale of the covariance is long, the covariance localization method is superior, while for short length-scales, the local analysis method is superior. Using the appropriate method for the appropriate length-scales can dramatically improve the performance of parallel EnKF.

*email: jeffsteward@gmail.com

*Preference: **Oral **