| 摘要 |
A parallel processing system capable of facilitating the data prefetch in the parallel computer and realizing a compilation scheme capable of removing the overhead caused by the control codes due to the parallelism. In the system, data dependency of each operation in the program is analyzed. Then, according to the data dependency, each processor recognizes those processors which have possibilities to require data allocated to the distributed part of the distributed shared memory in each processor, and each processor transmits the data to these processors before these processors actually require the data. The system has a host side compiler for compiling each program such that parallel processing independent portions of the program are compiled up to object-codes, while leaving parallel processing dependent portions of the program as intermediate-codes, and a node side compiler at each processor for optimally compiling the object-codes and the intermediate-codes obtained by the host side compiler to obtain optimum execution codes, according to information concerning parallelism.
|
