I/O Performance of a Large Atmospheric Model using PVFS
Abstract
Atmospheric models usually demand high processing power and generate large amounts of data. As the degree of parallelism grows, the I/O operations may become the major impacting factor of their performance. This paper evaluates the Ocean-Land-Atmosphere Model (OLAM) on the PVFS file sys- tem in order to point the I/O characteristics of the application. We show that storing the files on PVFS has lower performance than using the local disks of the cluster nodes due to file creation and network concurrency. Additionally, we study the performance of a new version of OLAM that used MPI associ- ated with OpenMP and show that the combined strategy presents I/O times 20 times shorter than the original MPI-only version and 9 times shorter on total execution time.