MULTICOMPUTER EFFICIENCY FOR EVOLUTIONARY PROBLEMS IN COMPUTA TIONAL PHYSICS

I.D. Sоfгоnоv, O.I. Sоfгоnоva
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 3-10.

      Currently the highest performance computer systems are represented by multicomputers that is multiprocessor systems containing a large number of processing elements interconnected by a switch network. The present paper shows that the multicomputer efficiency may be much lower than its maximum performance for problems which are well suitable for parallelization. One of the factors leading to losses in efficiency results from the necessity to implement a large number of transfers between processing elements which is due to the fact that the main memory is not shared by all processors.
      Some examples are given to see how the switch network architecture influences both the computer system performance degradation and the amount of hardware needed to achieve the desired speed.

A.V. Gusev, A.E. Lutsкy, I.L. Petrushсhenkov
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 11-14.

      The paper describes an approach to using algorithms for calculations in applied aerogasdynamics on multiprocessors. The parallelization of algorithms is exemplified by calculating the supersonic flow over a real aircraft.

E.L. Shnepоv, V.A. Pervоv, V.B. Glаzкоv
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 15-18.

      This paper presents a computer interconnection using fiber optics links for computer switching by means of the ES-4061 serial device via the IO interface of an ES computer.
      The structure and interaction concepts of communication hardware are described. The communication channel specifications are given.
The hardware described allows the intercomputer exchange rate to approach the highest throughput of the 10 channel while increasing considerably the distance between the computers without changing the operational principles of the serial ESr4061.
      The implementation based on the above hardware can be used as a generic component for creating local computer networks.

V.N. Ogibin, V.S. Khоushкin
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 19-27.

      A 2-D geometry definition language is considered. Specification methods and definition formats are presented for various 2-D geometry objects.

Yu.A. Bоndarenкo
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 28-36.

      For time-symmetric five- and seven-layer difference schemes with self-consistent operators, sufficient stability conditions are obtained, while for switching difference operators, sufficient and necessary conditions are derived such that the scheme remains stable when a low viscosity is added. Five- and seven-layer difference schemes are described which approximate the equation d²x/dt²+Ax = 0 with fourth order or higher accuracy and stability conditions are obtained. It should be noted that no completely stable schemes were found in the class under consideration. Several methods for deriving the discrete action functional are described such that its stationary-stable conditions yield preset accuracy difference schemes.
      The analysis was performed for a constant timestep and time-independent linear difference operators.

A.E. Кiуatкin, A.N. Вагуshniкоv
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 37-41.

      A method is presented for dividing the dynamic force characteristic of a coupling element in a one-degree nonlinear mechanical system into rigid and damping components. The coupling parameters were identified using the invariant immersion technique allowing to estimate simultaneously both state vectors and system parameters.
      The method capabilities are exemplified by defining force characteristics of two elements, one of them representing a sufficiently complex system where the paste flowing through the adjustment hole was used to create the stress.
      Theoretical oscillograms for element overloading and deformation agree well with experimental ones obtained from shock loading.

V.N. Pisarev
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 42-48.

      The parametric "Romb" scheme family is considered for a two-dimensional heat conduction equation. For some scheme parameter and difference mesh restrictions, approximation uncertainties, stability and convergence are investigated. The approximation Is studied for a heat conduction equation written in curvilinear coordinates in terms of normal heat flux components. The energy inequality method is used to prove stability and covergence. Equivalent schemes for temperatures are obtained on a diamond mesh. It is found that there is no schemes similar to those developed in this paper among well-known nine-point schemes.

I.M. Belyaкоv, O.V. Butneva, V.Ya. Bukharova, F.N. Vanin, I A.Gurenkova, T.V. Zabelina, I.N. Zvenigorodskaya, N.I. Kоzlоva, Yu.M. Matveev, A.N. Mоsкvin, A.V. Niкifогоva, N.T. Pletenyova, E.A. Ryazanova, V.V. Rуzhiкh, S.P. Santa1оv, V.A. Tarasov, L.P. Fedotova, R.M. Shagaliey, V.A. Shumilin
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 49-51.

      We consider SATURN program package designed for spectral transport calculations in two-dimensional cylindrical geometry The whole problem is divided into several computational regions where subgrids are created in order to construct efficient spatial grids. The package organization provides for modeling particle transport and interaction with medium in various approximations for each computational region while exchanging inner boundary conditions.
It is also possible to set different numbers of energy groups for each region and use a specific angular grid for each group Functionally, SATURN is divided into the following parts: command language translator, instruction set, initial data setting, computation control, operational modules.

V.V. Kоmоsко
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 52-58.

      The paper, which continues the publication on INTEDA database management system, describes a data model for computational physics problems, data access layers, a mechanism for interaction between the application program and ODDA and data manipulation language for ODDA.
      The data model in computational physics problems is a collection of database models. A database model is represented by a specific set of entities, each of them corresponding to some application domain notion, and by the relations between them. An entity may be internally complex, for example, it may be a multidimensional array, sparse matrix, linear or symmetric list etc.
      Two data access levels should be pointed out. The first level supports the access to entities considered to be a single element and the second one allows the access to separate items of the internal entity data structure. The first level is implemented using ODDA data manipulation procedures, while the other is formed either by using standard programming language tools or by creating a special package for each new data type.

В.M. Zhоgоv, B.A. Klopov, E.E. Meshкоv, V.M. Pasternak, A.I. Tоlshmуaкоv
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 59-65.

      Principles for grid selection and difference scheme creation are presented for two-dimensional gasdynamic calculations on irregular triangular grids. Local restructuring methods and reconfiguration introduction criteria are described. Physical experiment data and numerical results are given for a planar shock impacting the "heavy" angle.

V.T. Zhukov, O.B. Feodoritova
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 66-71.

      We present a finite-difference algorithm for solving the two-dimensional heat-conduct ion equation on parallel processors with distributed memory. To implement a code using the present algorithm, the spatial domain is divided into subdomains and each processor is assigned a single subdomain. The temporal difference equation is based on the explicit-iteration scheme with Chebyshev parameters. This algorithm has led to a very efficient parallel Fortran code on a MIMD parallel computer.

V.S. Kholushkin
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 72-76.

      A system for data conversion in a non homogeneous computer facility is described. Programming tools of the conversion system are used to operate various data types in program packages implemented on machines within a non homogeneous computer facility.

A.O. Lacis
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 77-80.

      An approach to the implementation of virtual common memory shared by multiple processors in quasi-transputer systems is presented. The approach is based on the concept of free migration of pages to the places where they are more frequently used. This optimizes automatically the data distribution between processors during Hun-time. Besides, any single processor gains access to large amount of virtual memory. Basic principles of the experimental implementation for the transputer system with linear topology are described. Relative performance considerations of traditional message-passing discipline and the approach presented are briefly examined.

V.P. Gorelov, A.N. Grebennikov, G.G. Farafоntоv, V.F. Yudintsev
VANT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 81-85.

      Calculations of Kef are used as an example to present a speed-up method for a direct procedure convergence when determining neutron transport equation eigenvalues with consecutive approximations.
      This method was analytically motivated for an infinite homogeneous medium using a single-group transport approximation. Muftigroup calculations were performed for spheres consisting of fissile material while varying the amounts of neutron moderator to illustrate its efficiency in finite geometries.

E.P. Shurina. Yu.G. Sоlоveiсhik, M E. Rоуak
NT. Ser.: Mat. Mod. Fiz. Proc. 1992. No 3. P. 86-87.

      Computer simulation is described for calculations of heat and electromagnetic fields in 3-D composite geometrically nonuniform structures.