Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Computational Science, Engineering & Technology Series
TRENDS IN PARALLEL, DISTRIBUTED, GRID AND CLOUD COMPUTING FOR ENGINEERING
Edited by: P. Iványi, B.H.V. Topping
The Scalable GPGPU-based Hybrid Computing System and its Applications
Keldysh Institute of Applied Mathematics, Russian Academy of Sciences, Moscow, Russia
B.N. Chetverushkin, "The Scalable GPGPU-based Hybrid Computing System and its Applications", in P. Iványi, B.H.V. Topping, (Editors), "Trends in Parallel, Distributed, Grid and Cloud Computing for Engineering", Saxe-Coburg Publications, Stirlingshire, UK, Chapter 7, pp 159-176, 2011. doi:10.4203/csets.27.7
Keywords: GPGPU, hybrid computing system, quasi-gas dynamic equations, gamma radiation transport, explicit schemes, parabolic equations, stability condition.
The need for new high performance computer (HPC) architectures became urgent during recent years. The reason of this is well known: the traditional architectures suffer from an increasing number of system bottlenecks. The number of transistors on the chips available today are evidently too large to be organized efficiently as in a traditional processor. The new HPC architectures are potentially very efficient, but good programming for them is often not easy for the application developers. The partial solution to this problem can be found by making the new HPC systems hybrid. The backbone of the system is just a well-known massively parallel processing (MPP) system, but each compute node of the system is additionally equipped with a co-processor with a non-traditional architecture. This hybrid approach makes porting of the applications from the traditional MPPs much easier, but it requires better inter-node communication systems. Possible solutions for the problems mentioned above are demonstrated on the MVS-Express supercomputer designed in Keldysh Institute of Applied Mathematics (KIAM). The compute nodes of the system are hybrid, each node consisting of the eight-core commodity server and a GPGPU as a co-processor. The inter-node communication system is also designed at the KIAM. It is a direct PCI-Express switch, highly optimized for latency to make hybrid application developments as easy as possible.
Hybrid computing systems require software to be created to take into account the hybrid structure of the memory. In this regard, explicit schemes which can be easily adapted to the architecture of the multicore systems including graphical processing unit (GPU)-based computers are very promising. But usually the explicit schemes impose severe stability limitations on the size of the time step, especially when the parabolic equations are solved. So development of explicit schemes with a rather mild stability condition is one of the most important trends. The original schemes are based on quasi gas dynamic (QGD) equation systems which describe viscous heat conducting flows which are good schemes for Navier-Stokes equations, where the latter are applicable. The original physically founded approach is proposed which permits Courant stability condition (tau~h) for explicit schemes to be achieved even for very low Mach numbers. The results of the numerical simulation of some CFD problems are presented. A good speedup in comparison with the universal processors was achieved. The use of an explicit scheme for the simulation of essentially subsonic flow of low compressible gas seems to be quite successful.
The technique of modeling gamma radiation transport in non-homogeneous structures on a HPC is also presented. Statistical algorithms based on "weighted" modifications of the Monte Carlo method are developed for modeling of gamma radiation in matter. Realization of developed methods on HPC MVS-Express is described.
purchase the full-text of this chapter (price £20)