News 
-
National Supercomputing Forum (NSCF-2014) was held on November 25th - 27th, 2014 in Russia, Pereslavl, the Program Systems Institute of RAS.
Presentations
Study the performance of the DVMH-programs on clusters with multicore processors and accelerators
Study the performance of the DVMH-programs on clusters with multicore processors and accelerators
The report analyzes the performance of Fortran DVMH (FDVMH) programs on different architecture clusters (using universal multcore processors, NVIDIA graphics accelerators and Intel Xeon Phi coprocessors in their nodes) . The characteristics of the programs written on FDVMH and their implementations on other languages are compared. The effect of various possible optimizations on the performance of DVMH-programs when mapping on the cluster is studied.
Presented by V.A. Bakhtin.
-
Scientific service on the Internet: the variety of supercomputing worlds was held on Semptember 22nd - 27th, 2014 in Russia, Novorossiysk, Abrau-Durso, a boarding house 'Sailor'.
Presentations
Automatic mapping of Fortran programs on clusters with accelerators
Automatic mapping of Fortran programs on clusters with accelerators
The paper presents the results of applying the system SAPFOR to parallelize sequential programs including programs with regular data dependences on clusters with graphical processing units. The system translates a program on the Fortran programming language in a program on Fortran DVMH. The obtained program is executed on a cluster. The Fortran DVMH language with compilers are the part of DVM-system. The paper proposes performance data of programs execution on GPUs and CPUs of the K100 cluster.
Presented by V.A. Bakhtin.
-
Parallel Computational Technologies (PCT) 2014 was held on March 31st - April 4th, 2014 in Russia, Rostov-on-Don, Southern Federal University.
Presentations
GPU parallelization of the tests from NAS NPB3.3.1 using Fortran DVMH programming language
GPU parallelization of the tests from NAS NPB3.3.1 using Fortran DVMH programming language
The paper presents a number of transformations applyed to the sequential NAS Parallel Benchmarks (NPB) of the 3.3.1 version (EP, MG, BT, LU, SP) and the parallel execution specifications of these tests by the DVMH directives which are essential for their high performance execution on clusters with GPUs. We explore the impact of different parallelization options on the programs performance. The performance of the tests developed on a high-level language Fortran-DVMH (hereinafter FDVMH) are compared with their implementation on a low-level language OpenCL performed by researchers from Seoul National University.
Presented by V.A. Bakhtin.