One of major ways to reduce the errors in numerical weather prediction is the improvement of the atmospheric state estimate. This estimate is used as initial data for a prognostic model. The computational complexity of the data assimilation algorithms dictates the necessity for efficient parallel implementation. In this work, a parallel implementation of the local ensemble transform Kalman filter data assimilation scheme is discussed. The results for the speedup and efficiency of the parallel implementation are presented. It is shown that a non-uniform distribution of the parallel workload between the processes restricts the efficiency of the parallel algorithm. Approaches to solving this problem are suggested.