Gpsro Data Processing Algorithm With Gpus

Unlike the conventional and satellite observations, the Global Positioning System (GPS) radio occultation (RO) techniques provide all-weather, high-vertical-resolution and global uniform coverage observations that require no calibration. Assimilating the GPS RO data into the global numerical weather prediciton system (NWPs) has a great siginificance for improving the accuracy of NWP. Currently, the GPS RO data is used in the following two ways: (1) retrieving the basic atmospheric parameters from the GPS RO data, and then assimilating these parameters into variational assimilation system of NWPs; (2) directly assimilating the bending angle and refractive index profiles in variational assimilation system of NWPs. Compared with the retrieval method, the latter method can provide higher accuracy. By using a forward model (or forwad operator) information in the space of model (i.e. NWP forecast) variavles can be mapped into that of the observations and back in a consistent manner.The forward model and the adjoint model are performed once a loop of the assimilation system which uses bending angle or refractivity. The method of bending angle needs most computing time, but it has the best error characters. For the forward model, there are about 4000 effective occultation events everyday, which need huge computation power to deal with. Most of researchers devoted to the parallel computing of GPS RO data process by using MPI and gained high speedup and efficency as well as scalability. But MPI can only exploit parallelism in process level, and can not make use of the fine grained parallel characteristics in larger scale of vector and Matrix operation.Recently, GPU's high computing capability, low power consumption and high cost-effective characters bring more and more people's attention. The GPS RO process contains immense data parallelism which suits the GPU very well. Therefore, we are motivated to accelerate the GPS RO data process on GPUs.We investigate the GPS RO forword model on GPUs and the main work is as follows:(1) The NVIDIA GPU architecture and the GPS RO observation principle are deeply investigated;(2) The parallel strategies for forward model of the GPS RO process are investigated and a preliminary GPU version is implemented; 600 profiles are processed on a single GPU and a cluster consisting 3 CPUs and GPUs, which obtains 205X and 549X speedup, respectively;(3) Many strategies, such as kernel splitting and fussion, improving the SM occupancy, eliminating the branch divergence and asynchronous execution, are utilized to optimize the performance, and the computing time reduced by 29%.