Maximize parallelism and minimize communication overheads are important issues for distributed memory systems. Communication and data redistribution cannot be avoided even when considering global optimization of data distribution and computation decomposition. A new approach based on loop fusion is presented exploiting pipelining parallelism, thus communication overhead can be hidden and data redistribution can be avoided. This technique exploits pipelining from complex loop structures, which distinguishes itself from traditional pipelining techniques. Ex-periments show that the technique is superior to other optimizations.
展开▼
