Extension of the DVM-model of parallel programming for clusters with heterogeneous nodes

Authors

V.A. Bakhtin, M.S. Klinov, V.A. Krukov, N.V. Podderugina, M.N. Pritula, Y.L. Sazanov

Annotation

The article discusses the principles of DVM-model extension  and Fortran DVMH language development for clusters with heterogeneous nodes. New  features  of the language allow to:   identify the parts of the program that should be run on a particular accelerator; to determine the data required for execution of a given program fragment; set of mapping rules of the loop iterations on the accelerator;  manage the movement of data between RAM  of universal CPU and accelerator memory. The principles of the Fortran DVMH compiler building and new functions of runtime system Lib-DVMH to support parallel program execution are described. Experimental data on the performance of the test program execution on GPUs of  K-100 cluster. The appearance of the Fortran DVMH compiler not only simplify the development of programs for clusters with heterogeneous nodes, but also will accelerate the creation for  such clusters  of automatically parallization compiler  from  Fortran language, that uses the Fortran DVMH as the output language on which the programmer can perform additional manual optimization of the program.

Key words

Parallel programming languages, hybrid multiprocessor computing systems, graphics accelerators.

Language

Russian

Reference

V.A. Bakhtin, M.S. Klinov, V.A. Krukov, N.V. Podderugina, M.N. Pritula, Y.L. Sazanov. Extension of the DVM-model of parallel programming for clusters with heterogeneous nodes // Vestnik of SUSU, series "Mathematical modeling and programming", Chelyabinsk: publishing center SUSU, 2012, No. №18 (277), Issue 12, P. 82-92