Calculation and Data Automatic Decomposition Method Based on Linear Algebra

In parallel recognition technology for distributed memory architecture, the key to improving parallel program performance is how to properly decompose computation and data to increase localization of data references and reduce communication between processors. Automatic data distribution requires simultaneous consideration of a series of issues such as program parallelism, locality, target machine characteristics, and back-end compiler optimization. This greatly increases the complexity and difficulty of automatic data distribution. The automatic data distribution framework using integer linear programming method can have better adaptability to target programs, but previous schemes have the drawbacks of high complexity and imprecise performance models. We propose a new highly practical automatic data distribution framework to address these issues. This framework provides support for multi-dimensional distribution, multi partition distribution, multi-layer pipeline parallel mining, and dynamic redistribution. It truly has high performance, scalability, and low overhead, thus having high practicality.