FEAST is a software package designed to solve large scale Finite Element based problems in a highly efficient way across a wide variety of modern hardware architectures. Core features and design goals are:
 |
FEAST is designed with both vertical and horizontal efficiency in mind. Horizontal efficiency means good scalability across a wide variety of MPI implementations and cluster computers. Vertical efficiency targets single-node performance with platform-specific Sparse Banded BLAS implementations for good memory performance and cache efficiency. FEAST also exploits the tremendous floating point performance and memory bandwidth of modern graphics processors (GPUs) as numerical co-processors. |
 |
It is completely concealed from both the user and the application programmer whether a FEAST application runs in parallel or serial mode. In parallel mode, the user does not have to care about partitioning, load balancing and communication. |
 |
FEAST encapsulates special adaptions for different architectures like commodity clusters, vector machines or other shared-memory supercomputers. This is a direct consequence of the core FEAST macro-wise domain decomposition scheme and its resulting local data structures. The user just compiles FEAST for the desired target architecture with MPI and does not have to care about manual adaptions. |