The high performance of supercomputers not only allows one to solve today problems faster but also allows one to formulate other problems, earlier not impossible to solve. However, this requires a revision of a programming model because of an increasing amount of the resulting data. MPI being a de facto a standard for distributed memory computing systems allows many options to actually program parallel I/O operations since the publication of the MPI-2 standard definition. However, such capabilities are often neglected when considering the design of applications for simulating a large problem. In this paper we describe some results of practical usage of MPI I/O in the context of gas filtration simulations through a porous mine medium. We also discuss the challenges presented by the scalability of application and performance of file subsystems and also describe how we perform a single process optimization and a parallel code optimization. All tests were run on a SKIF-Siberia computing system, a 566 CPU cluster located at Tomsk State University. This work was supported by the Russian Foundation for Basic Research (project N 08-08-12029-ofi).

Keywords: high-performance computing, hydrodynamics, filtration, parallel I/O, MPI-IO, scalability, data management

• PostScript (in Russian) (483KB)
• PDF (in Russian) (197KB)



• PostScript. zip (in Russian) (174KB)
• PDF. zip (in Russian) (112KB)