Improving Load Balance in a Weather Code: Asynchronous Output in HIRLAM with MPI

HIRLAM is a parallelized weather forecast software package. The standard HIRLAM suffers from load imbalance caused by the fact that one of the computational tasks handles the packing and writing of the intermediate results to disk while the other tasks are idle. We solve the load balance problem by assigning the packing and disk writing to an extra task and eliminating all the synchronization in the output. We discuss an MPI implementation of the asynchronous output scheme in detail and analyze its performance on an SGI Origin 2000, an IBM SP and an IBM p690.