International Conference «Mathematical and Informational Technologies, MIT-2013»
(X Conference «Computational and Informational Technologies for Science,
Engineering and Education»)

Vrnjacka Banja, Serbia, September, 5–8, 2013

Budva, Montenegro, September, 9-14, 2013

Glinskiy B.   Karavaev D.   Kulikov I.M.   Kuchin N.   Snytnikov N.  

High scalable computing using hybrid architecture supercomputer

Reporter: Glinskiy B.

Capabilities of the hybrid cluster for scalable computing of some applications are considered in article. Cluster is located in the Siberian Supercomputer Center and consists of 40 HP SL390s servers with three GPU NVIDIA Tesla M2090 on each server. Its peak performance is 85 TFlops including 79 TFlops on GPU’s. Two parallel computing models are used to achieve parallelism: SIMD on CUDA, according of GPU architecture, and MIMD on MPI.
The basis of the parallel implementation of solving the problems of continuum mechanics is a 3D domain decomposition, implemented using a MPI  and CUDA combination. The effectiveness of hybrid decomposition must becomputed only independently of flows across cell borders, for example, using Godunov type methods. The algorithm has potentially infinite scalability.
Parallel algorithm and program for 3D seismic field numerical modeling in inhomogeneous elastic media on hybrid cluster were developed. Numerical modeling is carried out on the basis of difference method. In program GPU is used to implement the computational part of the algorithm on CUDA, CPU is used to exchange data between the GPU with using MPI. More than 30 thousand cores were used in calculations.
Parallel algorithm for a solving of Poisson equation is based on a 3D domain decomposition, partial pre-computation of Green's function and calculation of a single-layer potential. It is implemented using MPI andaimed to run on a CPU subsystem of a hybrid cluster. To solve Vlasov equation we use particle-in-cell method, implemented for a GPU subsystem with the use of CUDA. Developed algorithm allows to conduct numericalsimulations with billions of grid cells and hundred billions of particles.
Technological approaches for scalable computing on hybrid clusters with GPU are discussed.
 

Full text file: GlinskiyBM.pdf


To reports list

© 1996-2019, Institute of computational technologies of SB RAS, Novosibirsk