High-performance land surface modeling with a Linux cluster

The Land Information System (LIS) was developed at NASA to perform global land surface simulations at a resolution of 1-km or finer in real time. Such unprecedented scales and intensity pose many computational challenges. In this article, we demonstrate some of our approaches in high-performance computing with a Linux cluster to meet these challenges and reach our performance goals. These approaches include job partition and a job management system for parallel processing on the cluster, high-performance parallel input/output based on GrADS-DODS (GDS) servers, dynamic load-balancing and distributed data storage techniques, and highly scalable data replication with peer-to-peer (P2P) technology. These techniques work coherently to provide a high-performance land surface modeling system featuring fault tolerance, optimal resource utilization, and high scalability. Examples are given with LIS's high-resolution modeling of surface runoff during 2003 to illustrate LIS's capability to enable new scientific explorations.