For many years, 4DWX ran operationally on geographically distributed HPCs at the ATEC ranges, each running its own NWP model using local dedicated RAID storage. With more recent advances in HPC technology, it has become feasible to run multiple NWP models on centralized HPCs using shared network storage. This approach provides for much greater levels of efficiency, flexibility, and extensibility at only an incremental cost.
NCAR's experiences evaluating, designing, implementing, and transitioning from a distributed to centralized architecture will be described as a case study. The scientific requirements included forecast lengths, grid sizes, number of nested domains, and post-processing products. Some of the engineering considerations were connectivity between compute nodes, memory management, file transfer and management, job queuing and node assignment, system security, and distribution of model outputs to analysis and display systems. Pitfalls as well as successes will be summarized and thoughts about the future of HPCs for fielding operational NWP systems such as 4DWX will be offered.