Automation of parallel program development for clusters with accelerators


V.A. Bakhtin, N.A. Kataev, A.S. Kolganov, V.A. Krukov, N.V. Podderugina


The main difficulty in the development of parallel programs for the cluster is the need to make global decisions on the distribution of data and calculations taking into account the properties of the whole program, and then perform painstaking work on program modification and debugging. A large amount of code, multi-modularity, multifunctionality makes it difficult to make decisions on the coordinated distribution of data and calculations. To solve this problem the method of incremental or partial parallelization can be used. The idea of this method is that not the whole program is subjected to parallelization, but its parts (areas of parallelization) – additional copies  of the required data are created in them, the distribution of the data and the corresponding calculations are performed. These areas can be built based on the times obtained by profiling a sequential program.

Key words

DVM, program parallelization, distributed memory, incremental parallelization, luster.




The paper is not published yet.