Intellectual Personal Supercomputer for Solving R&D Problems

ISSN 2409-9066. Sci. innov. 2016, 12(5): 15—27 doi: https://doi.org/10.15407/scine12.05.015

Khimich, O.M., Molchanov, I.M., Mova, V.І., Nikolaichuk, О.O., Popov, O.V., Chistjakova, Т.V., Jakovlev, M.F., Tulchinsky, V.G., and Yushchenko, R.А. Glushkov Institute of Cybernetics, the NAS of Ukraine, 40, Glushkov Av., Кyiv, 03187, Ukraine, tel.: +38 (044) 526-60-88, fax: +38 (044) 526-41-78 INTELLECTUAL PERSONAL SUPERCOMPUTER FOR SOLVING R&D PROBLEMS

New Ukrainian intelligent personal supercomputer of hybrid architecture «Inparcom-pg» has been developed for modeling in defense industry, engineering, construction, etc. Intelligent software for automated analysis and numeric solution of computational problems with approximate data has been developed. It has been used for implementation of simulation applications in construction, welding, and underground filtration. Keywords: modeling, simulation, intelligent computer, hybrid architecture, computational mathematics, and approximate data.

Mathematical modeling and associated com- computer capabilities despite the multi-core pro- puter experiment is currently among the main cessors. GPGPU technology (from General-Pur- tools for studying various natural phenomena pose Graphic Processing Units) meets the de- and processes in society, economy, science and mand of accelerating calculations on multi-core technology. Simulation significantly reduces the computers for big numbers of similar arithmetic development time and cost of new objects in en- operations. It means general purpose computing ergy and resource-saving. It increases efficiency on video cards. of real test planning and makes it possible to con- The approach has bumped the development of sider several variants to select the best. specialized graphic processors. They implement Nowadays the computer performance growth the SIMD parallel computing architecture. SIMD is achieved due to parallel calculations based on means the architecture with single command use of multiple processor devices and multi-core stream and multiple streams of data. The short- processors. In late 2014, Intel announced a pro- term future of high-performance computing hard- cessor series with the number of cores varying ware seems to be related to the hybrid systems from 4 to 18, and then in early 2016, from 4 to 22. combining MIMD and SIMD architectures toge- Note the processors implement MIMD-archi- ther, such as multi-core computers with GPGPU tecture (the architecture with multiple streams accelerators. of both commands and data). At the same time, the requirements for high-performance comput- MOTIVATION FOR THE PERSONAL INTELLIGENT ing are far ahead of the conventional parallel SUPERCOMPUTER DEVELOPMENT Application of the latest GPGPUs makes it © KHIMICH, O.M., MOLCHANOV, I.M., MOVA, V.І., possible to achieve up to 2.62 TFLOPS of double NIKOLAICHUK, О.O., POPOV, O.V., CHISTJAKOVA, Т.V., JAKOVLEV, M.F., TULCHINSKY, V.G., precision calculations in a usual computer system and YUSHCHENKO, R.А., 2016 unit. This is a big incentive to implement a «per- 15 Khimich, O.M., Molchanov, I.M., Mova, V.І., et al. sonal supercomputer» for high performance com- of Ukraine together with Electronmash have de- puting of individual use. veloped a concept and an experimental model of Under conditions of fundamentally new 3D «Inparcom-pg» personal intelligent supercom- models, transition from the by unit computer puter of hybrid architecture for solving the scien- modeling to the whole product simulation and tific and engineering problems. complete optimization the significant improve- ment of the modeling and design quality is pos- CONCEPT OF PERSONAL sible in defense industry, engineering, construc- INTELLIGENT SUPERCOMPUTER tion, etc. The innovation of the personal intelligent su- A character feature of application mathemati- percomputer development and use consists in cal models is the need to account their input da- implementation of three major modeling para- ta inaccuracy together with the model equations digms: computational mathematics, high-perfor- [1—4]. A numeric simulation key issue which mance computing and intelligent software. The accumulates the impact of all other factors is the paradigm implementation makes it possible to computer solution validity. The issue presence shift the work on formulating and solving prob- and importance is supported even by the fact of lems from the user toward computer in compari- NAFEMS (National Agency Finite Element Me- son with traditional computer technologies, to thods and Standards, UK) groups working over reduce the applications development time need- twenty years in 30 countries to ensure reliabi- ed to solve scientific and technological problems, lity and safety of the engineering calculations and to improve the quality of computer solu- based on finite element method and related tech- tions [10]. niques. Computer models of application problems al- For parallel computers, the above mentioned ways contain inaccuracy in input data. However, issue is supplemented by the demand for ac- the vast majority of existing software used in ap- counting the computer architecture and techni- plication problem studying implicitly assumes cal parameters, prediction of the optimal num- the initial data be precise. The characteristic of ber of processes, planning data distribution among computer models with approximate data are a processes, implementation of synchronization of priori unknown mathematical features. Within a calculations and data exchanges, etc. [5—7]. So given level of inaccuracy one can find both con- the development of parallel algorithms and appli- sistent and inconsistent problems, both well-po- cation programs requires much time, and their sed and ill-posed ones, both well-conditioned and successful usage depends on a specific high qua- ill-conditioned ones. In this context a computer lification of users. The issues of obtaining reli- problem which is eventually solved is always an able computer solution, of decreasing the time approximation of the original problem (because spend for scientific or engineering problem de- of the inherited input data inaccuracy and/or scription can be solved by intelligent computers sampling error and/or numeric representation [8—10] through involvement of the computers error) [1, 5, 7]. to studying the model features and generating The situation is strongly complicated by the the parallel programs. fact that big mathematical difference between the As a part of the innovative R&D project «Deve- matrices of complete and incomplete rank exists lopment of Hardware-Software System Based on only in the ideal world of mathematical real num- the Personal Intelligent Supercomputer of Hy- bers. Since computer operations on the matrices brid Architecture for Mathematical Modeling in are executed with rounding-off, the difference Defense, Engineering, and Construction Indus- becomes uncertain. Thus, an invertible matrix tries» Glushkov Institute of Cybernetics of NAS can become degenerate in a computer. On the 16 ISSN 2409-9066. Sci. innov. 2016, 12(5) Intellectual Personal Supercomputer for Solving R&D Problems other hand, a degenerate matrix is likely to be computing virtualization with generic support turned into a close non-degenerate one due to the of multiple operational systems, system moni- rounding errors. toring, connection to grid infrastructure. Analysis of computer arithmetic implementa- High performance of intelligent supercompu- tion has shown its difference from the conven- ter packed in a personal computer case (up to 3— tional arithmetic. It is related both to the round- 9 TFLOPS) is achieved by combining software ed representation of real numbers itself and to and hardware dense computing tools, the intelli- the simulation of associative low, of distributive gent computational problem solver and the appli- low, of commutative low, and so on for the arith- cation software virtualization. Besides, for analy- metic operations. I. e., the mathematical axioma- sis and solution of basic computational mathe- tic of computing differs from the computer arith- matic problems with approximate data including metic axiomatic. those for sparse data structures performance is The issue resolving implies detection of the com- increased through use of intelligent and specia- puter problem features and generation of machine lized algorithms. Intelligent personal supercom- algorithm to obtain an approximate solution of puters occupy a niche between supercomputers the mathematical problem regardless of its well- and modern personal computers. posed or ill-posed, well-conditioned or ill-condi- The experimental model of «Inparcom-pg» in- tioned nature. telligent personal supercomputer implements the Therefore, for each class of mathematical prob- above concept. lems with approximate data discussed further a software toolkit to study the computer model ARCHITECTURE AND EQUIPMENT mathematical features and to solve the problem OF INTELLIGENT PERSONAL SUPERCOMPUTER has been developed. It selects the best fitted «Inparcom-pg» personal supercomputer is the computer algorithm and adjusts its parameters next step in development of «Inparcom» family according to the computer parameters and ar- of intelligent workstations designed by Glushkov chitecture. Finally it estimates the results vali- Institute of Cybernetics in cooperation with dity [3, 7]. Elec tronmash in 2005—2012. The architecture Thus, the main conceptual principles of the has shifted toward higher calculation density intelligent hybrid supercomputer architecture provided by combination of modern GPGPU ac- are the next: celerators, advanced numerical methods, new adaptive automation of the hybrid computer technologies of processing and storage of big in- effective topology building on the base of com- formation amounts, application virtualization soft- puter algorithms of mathematical feature ana- ware [11—13]. lysis for the problem computer models; «Inparcom-pg» is a knowledge-oriented com- guarantee of the solution accuracy require- puter which gains knowledge about the problem ments according the computational problem computer model features during solving an en- model by either software or hardware means gineering or scientific problem, automatically based on long number arithmetic and provid- builds an algorithm, a program and a topology of ing the result confidence rating; the hybrid computer correspondent to the dis- computing density due to the innovative hyb- covered features, and estimates the result reli- rid methods of sparse data processing which ability after finishing the computing. are adapted to the latest models of NVIDIA The intelligent hybrid personal supercomputer Tesla graphic accelerators based on Kepler ar- hardware (fig. 1) consists of: system unit, monitor, chitecture with the latest CUDA version, and keyboard, mouse, and uninterruptible power sup- Inparcom-pg general architecture; ply, printer and scanner (or MFP optional). ISSN 2409-9066. Sci. innov. 2016, 12(5) 17 Khimich, O.M., Molchanov, I.M., Mova, V.І., et al.

quad-core personal computer in double precision arithmetic. In other words, the time of solving several application problems is reduced from five days to one hour.

SYSTEM SOFTWARE CONFIGURATIONOF INTELLEGENT SUPERCOMPUTER System software configuration of «Inparcom- pg» personal supercomputer provides support for several modes of use related to the application software features: personal computer mode; remote desktop mode; Fig. 1. Components of «Inparcom-pg» remote computing service mode. Figure 3 shows the system software configuration including the operation systems, virtual machines, compilers and libraries. The system software basic level is presented by Linux operating system. It provides development and runtime environments with highest performance access to physical hardware. Virtual Ma- chines deployed over the Linux basic level create application environments familiar for end-users of each program with graphic user interface. All computational programs are executed in the Linux basic level under the System Monitor control to benefit of the level performance [14]. The fundamental idea of virtual machine technology is abstraction of computer hardware (processor, memory, disk, network, etc.) so to have multiple execution environments with illusion for each of the environments to be is performed on a separate physical hardware. Virtual machines enable complete system fitting to the application specificity and sharing the host ma- Fig. 2. System unit of «Inparcom-pg» chine resources on demand. Virtual machines access the host machine by virtual network and The system unit (fig. 2) adapted to scientific and thus can perform calculations on the host ma- engineering problems contains 2 to 4 GPGPUs chine. Besides, a user can work with application NVIDIA Tesla K40 with peak performance of software via the host operational system graphic 1.43 TFLOPS and 12 GB memory each, at least interface, and if necessary use the virtual ma- 64 GB RAM, and highspeed SSD. chine manager. A user can access applications As result «Inparcom-pg» personal supercom- remotely by VNC (for Linux) or Remote Desk- puter about 100 times outperforms conventional top (for Windows). 18 ISSN 2409-9066. Sci. innov. 2016, 12(5) Intellectual Personal Supercomputer for Solving R&D Problems

INTELLIGENT SOFTWARE FOR SOLVING SCIENTIFIC AND TECHNICAL Virtual network PROBLEMS Intelligent numerical software was developed to examine and solve basic problems of computational mathematics with approximate data as a Applications component of the considered system [15]. The intelligent software structure includes a library Inparlib_pg of intelligent programs for ana- System monitor lysis and solving problems of the next classes: sys- Compilers, libraries tems of linear algebraic equations, algebraic eigen- Inpartoolpg and services Windows value problems, systems of nonlinear equations, KVM Virtual systems of ordinary differential equations with Linux Machine Hypervisor boundary conditions. It also include Inpartool_pg intelligent software tool that implements the prob- Hardware lem description in the subject domain language, automatic execution of the computational mathe- Fig. 3. Configuration of «Inparcom-pg» system software matics problem analysis and solution processes with the results validity assessments. process per core). The process here means a program that runs in parallel with others, uses its part Inparlib_pg Intelligent of local memory and coordinates with other pro- Program Library cesses by information exchange implemented in Let call intelligent a program which solves a the form of send/receive operations. Parallel pro- problem by a specified algorithm and during exe- cesses of such type are implemented by the stan- cution checks the user-selected algorithm appli- dard for CPU computing MPI system. cability to the computer model features; auto- Each the process, in its turn, is able to paral- matically generates the efficient configuration lelize calculations on a number of GPU threads. of hybrid computer and fits the best number of Usually, a process uses one GPU. Within the de- central processor (CPU) cores and graphics co- velopment of a parallel algorithm one has to de- processors (GPU) to minimize the calculation ti- sign the task subdivision into subtasks i.e. to me; distributes the data among processors; sol- determine the subtask priorities; to select better ves the problem and makes confidence estimate computing resource distribution among the sub- for the output solution. It provides load balan- tasks; to specify which subtasks can be perfor- cing, synchronization of data exchange between med in parallel; to coordinate the sequence in CPU and GPU, minimization of communication which the subtasks should be performed, etc. overheads. Thus, the hybrid approach involves two levels of In terms of use, the Inparlib_pg library programs parallelism: parallelism among the process level are functional components of the Inpartool_pg subtasks and the intra-subtask parallelism by intelligent tools being it the same time indepen- GPU threads. dent reusable components for solving the appli- When implementing the CPU-level subtasks cation problems where computational mathema- it is necessary to take into account connections tics is used at intermediate or final stage. between processors and between cores within The developed hybrid algorithms and prog rams the processors. Much execution time is spent for are designed so that the algorithm execution is exchanges between the processes. Therefore, it is divided into processes run on CPU cores (one important to select the most efficient topology ISSN 2409-9066. Sci. innov. 2016, 12(5) 19 Khimich, O.M., Molchanov, I.M., Mova, V.І., et al. for a particular problem, to fit the optimal num- and software developers is prediction which sub- ber of CPUs and to specify interconnection be- tasks in each algorithm will be more efficiently tween them. MPI makes it possible to create vir- implemented on CPUs and which on GPU. tual to-pologies of fixed number of cores (processes). Virtual topology means the software to- Intelligent Inpartool_pg pology of relations between the processes such as Software Toolkit ring, grid, torus, hypercube, etc. regardless hard- Inpartool_pg is an intelligent software toolkit ware communication environment. The virtual of the end-user level. It provides user/computer topology model can be presented by a graph with interaction in the problem domain language and vertices of processes and edges of communica- executes automatically all stages of computer prob- tion channels. Thus, a MPI user (programmer) lem solving (algorithmic, programming with con- does not need to know the scheme of physical sidering approximate initial data and the com- links between the cores which greatly simplifies puter solution validity analysis) [15]. the process of writing a parallel program. How- Conceptual principles of process flow chart ever, the user has to determine the number of for problem solving in the intelligent interface cores for solving the task and distribute the da- include: ta between them ensuring their balanced load. Capability to solve problems with approxima- Since the op-erations of data transmission far te input data; exceed the dura-tion of arithmetic operations and Problem specification in the problem domain memory access operations, it is necessary to dis- language; tribute the data in the processor memory so to User-friendly methods of data input; provide good balance between the data exchan- Automatic program synthesis on the base of ge and arithmetic operations performed concur- computer studying of the problem mathemati- rently and minimize the whole time of executing cal features, knowledge of the system and the the program. When developing algorithms for problem to be solved; hybrid architecture it is necessary to take into Output of problem solution together with the account the differences in memory models for computer result validity assessment; CPU and GPU. CPU programs can address di- Problem solution supplemented by a protocol rectly any cell of linear and homogeneous me- of the process of studying the features and mory. Besides, modern com-puter processors ha- evaluation of the results; ve fairly big internal cache which implicitly im- Implicit parallelization. The implicit paralle- proves the performance of algorithms and pro- lization assumes [15]: grams on CPU. Automatic hybrid program subtask distribu- Peak performance of graphics processors great- tion among CPUs and GPUs; ly exceeds CPU performance. Yet GPU implemen- Automatic fitting the numbers of CPUs and tation of parallel computing based on NVIDIA GPUs for the program execution, configura- CUDA technology contain six types of memory, tion of the computer for the problem efficient each with own function and performance. The solution; main issues of parallel computing for hybrid ar- Automatic distribution of input data among chitecture are: need to master different tools for the program processes according the algo- CPU and GPU parallel programming; complexi- rithm; ty of resource distribution among the processor Load balancing; cores and graphics coprocessors; limitations of Synchronization of inter-process data ex- communication between CPU and GPU. In ad- change; dition, one of the important tasks of algorithm Coordination of asynchronous operations; 20 ISSN 2409-9066. Sci. innov. 2016, 12(5) Intellectual Personal Supercomputer for Solving R&D Problems

Minimization of inter-process data exchange. Web Browser User Interface As result, the user experience of Inparcom-pg Input Output personal supercomputer with both multi-core and graphics processors is similar to of a serial Remote Access System personal computer. for Linux &Windows Inpartool_pg implements the concept of knowledge driven development. Its design is based on synthesis of the major benefits of modular Authentication Service programming, databases, knowledge bases and the advanced knowledge engineering methods Authorization Service of information presentation, storing, acquisiti- oning, etc. Inparlib_pg Library Explanatory The intelligent software architecture was foun- of Program Modules System ded on formal models of the subject domain of Help System and online Manuals computational mathematics: the most common Planning and Control problems, methods and algorithms of their sol- System Knowledge base ving on hybrid computers. The formal model is closely connected to the principles of automa- Fig. 4. Client-server architecture of Inpartool_pg tic analysis of the computer model features in order to automatically generate the algorithm and to synthese hybrid program which accounts sults validity set the next requirements for the the input data approximate nature, mathemat- planning and control system: ics and technical characteristics of the hybrid information gain from the input data analysis computer. and its transformation to the initial knowledge Besides, the software design supports use of about the problem features and preferable al- Inpartool_pg through Internet. The software is gorithm; of client-server architecture. The client side con- application domain knowledge collection and sists of interactive frontend only. And the server processing during the computing planning; side includes systems that provide the users ac- transformation of the knowledge on the prob- cess to the intelligent software backend as well as lem features to the algorithm choice and the analysis and solution of the problems with ap- program synthesis; proximate data on the hybrid computer. Figure 4 automatic input data distribution among the schematically depicts the Inpartool_pg client- processes; server architecture. output and storing the solution results for fur- The remote access, authentication and autho- ther explanation and presentation. rization services provide user with both direct After the semantic graph based formal analysis and Internet access to Inpartool_pg. of the input data features the software modules The Planning and Control System is related to are selected for further analysis of the problem the problem domain specification (in the form of features. Then the problem features are automa- semantic network or graph), the knowledge base tically converted to the algorithm choice, and the and the user interface. Main objective of the com- program is synthesized. Within the synthesis the puting planning is to find the best method of the computer resources are selected, the virtual in- problem solving with the least user interaction. terconnection configuration is generated and the The principles of automated computer prob- computing process is being continued according lem analysis and solving with assessment of re- to the semantic graph. Both control and data ex- ISSN 2409-9066. Sci. innov. 2016, 12(5) 21 Khimich, O.M., Molchanov, I.M., Mova, V.І., et al.

Inparcom_pg (1 GPU), sec 57 PC, sec 3,254 = 54 min 14 sec

SLAE order 661,590 Band width 541,257 Density, % <1

SLAE order 283,031 Band width 19,530 Matrix density, % 7 Fig. 5. Results of SLE solutions in LIRA_g change communication channels between the in- of user training in use of the intelligent software: volved software modules are established. both automatic and interactive automated prob- The Explanatory System explains how the lem solving modes are supported. Besides, they problem solution was obtained and reasons why satisfy the requirements to communicate in the the particular feature analysis process selected. subject domain language and provide convenient It presents the problem solution with the confi- intuitive forms for information input/output to dence estimates or the reasons for inability to make unnecessary the paper documentation. obtain it. A user can control the level of detail of the explanations. There are different scenarios APPLICATION SOFTWARE of explanations and different level protocols of Numerous software tools (from programs to computing for this purpose. big packages) have been developed over the last The Help System provides a user with guides decades for solving a wide variety of application and manuals how to work with Inpartool_pg. problems. In the majority of the cases the tools The user interaction with Inpartool_pg is im- were designed for computers of traditional serial plemented via a system of dialog web forms, architecture. Adaptation of the tools for the hyb- namely: rid architecture computers requires their signifi- problem statement and data input; cant modernization by top level application soft- computing process control; ware developers. presentation of solution results; An example of such success story is the pro- explanatory; gram complex ANSYS [16]. It proposes solutions help and reference information. for a wide range of modeling problems in compu- The dialog scenarios have been developed on tational aero-hydrodynamics, solid mechanics, the base of the application domain model. They elect ro-magnetic simulation, etc. ANSYS deve- account different user goals and different levels lopers implemented (e. g., in ANSYS® v. 16.1) 22 ISSN 2409-9066. Sci. innov. 2016, 12(5) Intellectual Personal Supercomputer for Solving R&D Problems the aero-hydrodynamics parallel computing tools suitable for both workstations and distributed systems equipped by either multi-core processors or graphics coprocessors. Another approach to parallelization of existing serial software is just replacement of their computational libraries by the new parallel ones at least for the operations especially greedy for computing resources such as processing time, memory, etc. Intelligent software of kind Inparlib_pg library is the best choice for such replacement. It can be used for example to solve systems of linear algebraic equations (SLE). The precondition of Geometry of water saturated layers of Chernihiv this approach is full access to the serial software region with branched network of surface stream flows source code. Inparlib_pg was designed for Linux. (the model covered area is 184 × 222 km) However, if the original software runs under Solution time: Windows the Inparcom-pg technology proposes Inparcom_pg (1 GPU), sec 65 a solution with the application run on a Windows PC (8 cores), sec 1,266 = 21 min 6 sec virtual machine and the Inparlib_pg run in the SLEA size: host Linux environment. System order 1,151,112 Band halfwidth 5,367 Let’s consider three examples of such approach Memory, Gbyte 20 implementation. Fig. 6. Simulation of underground filtering in Nadra-3D_g Information Technology of Strength Analysis of Building Construction PC Lira_g users who aren’t IT professionals lack confidence The problems of strength analysis arise in dif- when send proprietary data via Internet to 3rd ferent sectors of the economy such as construc- party hardware. tion and mechanical engineering. Increased de- The personal supercomputer of hybrid archi- sign quality demands together with distribution tecture Inparcom-pg makes it possible to use of new construction technologies and materials high- performance computing locally. To meet generate new requirements to numeric model- users desires the new information technology of ing. In addition, architecture development stim- building constructions strength analysis Lira_g ulates the steady need in calculation of the com- was implemented as Inparcom-pg application. It plex unique objects. So the demand for new so- is organized so. The conventional LIRA run in a lid mechanics methods and computational ap- Windows virtual machine prepares data files for proaches able to provide the correct and precise parallel calculations. Then the data files are cop- computer modeling for real constructions is ied through virtual network in the host Linux growing. en-vironment where the intelligent software is The program complex LIRA [17] is developed immediately started to distribute the data be- for building constructions strength calculations tween the computing units and to execute cal- on personal computers. Its branch LIRA-cluster culations. The results are converted in the LIRA [18] developed to use computing clusters for format and copied back to the virtual machine, high er performance calculations. Its disadvan- and then LIRA catches them and presents. Fig. 5 tage is the need for users to copy big data volu- shows the effect of solving SLE of LIRA by In- mes to and from remote clusters. Besides, LIRA parcom-pg. ISSN 2409-9066. Sci. innov. 2016, 12(5) 23 Khimich, O.M., Molchanov, I.M., Mova, V.І., et al.

hiv region. The computational grids of tetrahe- dral finite elements built for 3D model of the region is highly detailed. It accounts the strata geometry and parameters together with the surface river network. The computational problem is reduced to SLE with wide band matrices. Nadra-3D ported to Inparcom-pg in the form of information technology Nadra-3D_g targets such problems. The system uses Inparlib_pg for reliable and quick solution of big SLE (from 400 thousand to 1.5 mil- lion variables, Fig. 6). The intelligent software use for SLE solving provided significant (tenfold) speedup of the Nadra-3D finite elements modeling. The order of solved SLE: 106—1012).

z Information Technology for Simulation of Welding Processes and Related Technologies Technical condition diagnostics of welded structures under significant strains or in aggressive environments is the key to their stability and da- r mage protection. Since welded structures are used s in nuclear and thermal power plants, pipelines o β β = const and other objects of high safety requirements, the δ α z = const most precision evaluation of the structure state β from the known operational da-mage such as the surface corrosion metal loss is important to en- r = D/2 sure their long-term reliability. E.O. Paton Electric Welding Institute of NAS SLEA order 55,560 of Ukraine has developed the software package Band half width 1,052 WeldPredictions for application research of ther- Fig. 7. Simulation of stress-strained state of weld structures mo-mechanical processes of welding and furth- in WeldPredictions_g er exploitation of welded structures. It consists of serial personal computer programs. 3D mo- Information Technology del processing by the programs is computati- Nadra-3D_g for Spatial Heat onally hard and requires up to several days of and Mass Transfer Modeling continuous calculations to get result. The Finite elements method is a popular approach to WeldPredictions_g information technology [20] numerical modeling of physical processes. Glush- has been developed for solving the processing kov Institute of Cybernetics has developed the problems on Inparcom-pg intelligent personal finite element solver Nadra-3D [19] for simula- supercomputer with Inparlib_pg library SLE tion of mass transfer processes such as under- program use (fig. 7). ground water filtering in large scale multi-com- WeldPredictions_g 20-60 times speedups the ponent geological environments tested in Cherni- processing. 24 ISSN 2409-9066. Sci. innov. 2016, 12(5) Intellectual Personal Supercomputer for Solving R&D Problems

CONCLUSIONS neering, nuclear power, aircraft and ship design- AND IMPLEMENTATION PERSPECTIVE ing, defense industry, civil and industrial con- The innovative scientific and technological pro- struction, electric welding, and economics. ject «Development of hardware-software system The experimental model of Inparcom-pg is based on intelligent personal supercomputer of used in Electronmash for scientific and engineer- hybrid architecture for mathematical modeling ing calculations including strength analysis of in defense industry, mechanical engineering and construction projects, underground filtering si- construction» resulted in implementation of hyb- mulation, modeling of stress-strained states of rid intelligent personal supercomputer Inparcom- welded constructions. pg including development of several algorithms Inparcom-pg application for the scientific and and programs for hybrid computers to solve prob- engineering problems significantly reduces the lems using the intelligent software which imple- time of problem solution and ensures the result ments the innovative functions of automatic adap- quality. Its design reduced the size of parallel tive selection of algorithm, program parameters computer to the personal computer format and and hardware configuration on the base of the decreased its power consumption through use of problem features. modern graphics accelerators. In addition, the approximate nature of input Inparcom-pg is a high performance computer data is accounted, and the solution validity is designed for personal use. It is important for the ensured. The algorithms and software for big applications where information protection is of sparse dataset processing typical for modeling high priority, e. g., for design automation in con- problems in defense, engineering and construc- struction. tion are developed. REFERENCES High performance of the personal computer (up to 3 TFLOPS) has been achieved by highly 1. Molchanov I.N. Machinye metody reshenija prikladnyh optimized computational libraries, advanced tech- sadach. Algebra, priblizhenie funkcij. Kyiv: Naukova dumka (Scientific thought), 1987 [in Russian]. nical means of hybrid architecture and virtualiza- 2. Khimich A.N. Оtsеnki vozmushchеnii dlja rеshеnija zа- tion of application software. dаchi naimen'shikh kvadratov. Kibernetika i sistemnyi The intelligent parallel supercomputer main analiz (Cybernetics and systems analysis). 1996. 3: 95— advantages are the next: 102 [in Russian]. 3. Khimich A.N. Оtsеnki pоlnоi pogreshnоsti reshеnija release of a user (programmer) from analysis of sistem lineinykh аlgebraichеskikh uravnenii dlja matrits problem features, development of algorithms proizvol'nogo ranga. Komp’jutеrnaja matematika (Com- and software for hybrid computers reduces the puter mathematics). 2002. 2: 41—49 [in Russian]. problem statement and solution time for sci- 4. Khimich A.N., Voitsekhоvsky С.А., Brusnikin V.N. О dostovernosti lineinykh matematicheskikh modelei s ence and engineering; priblizhonno sadannymi iskhodnymi dannymi. Ma te- use of the subject domain language for com- maticheskie mashiny i sistemy (Mathematical Machines puter specification of problems with approxi- and Systems). 2004. 3: 54—62 [in Russian]. mate data; 5. Mikhalevich V.S., Molchanov I.N, Sergienko I.V., Bik N.A., Burlaka E.N., Burchak O.T., Viznjuk G.I., Vino ku ro- the problem computer solution with confi- va I.P., Galba E.F., Geets E.G., Gerasimova T.A., Ger- dence estimates; shovich V.I., Gorlach S.P., Guljanitskii L.A., Zubaten- implicit paralleling provides the comfort of us- ko V.S., Kalenchuk-Parkhanova Zh.A., Kapitonova J.V., ing the parallel computer just like a conven- Kaspshchitskaja M.F., Klimashenko J.A., Kuksa A.I., Lev- tional serial computer. chenko I.S., Letichenskii A.A., Luchka A.J., Mazira G.P., Nezlina A.J., Nesterenko A.N., Nikolenko L.D., Nosh- Inparcom-pg intelligent personal supercompu- chen ko O.N., Orlova K.F., Platon E.V., Pogrebinskii S.B., ter of hybrid architecture can be used for a range Reshetukha I.V., Rudich O.V., Rjabtsev V.E ., Semen chen- of modeling problems in the areas such as engi- ko S.P., Stiranka A.I., Tukalevskaja N.I., Khimich A.N., ISSN 2409-9066. Sci. innov. 2016, 12(5) 25 Khimich, O.M., Molchanov, I.M., Mova, V.І., et al.

Khodzinskii F.I., Chernenko A.C., Chistjakova T.V., Shash- ber netika i sistemnyi analiz (Cybernetics and systems kina S.I., Shor N.Z., Jakovlev M.F. Chislennye metody dlja analysis). 2007. 3: 51—55 [in Russian]. mnogoprotsessornogo vychislitel’nogo kompleksa ES. Pod 14. Semerenko V.R ., Falfushinsky V.V., Yushchenko R .A. red. I.N. Molchanova. Moskva: VVIA im. prof. N.E. Zhu- Con struction of the job management system for family kovskogo (M: Air Force Academy named after Zhukov- Inparkom clusters. In Proc. 3th International conference sky professor), 1986 [in Russian]. «TAAPSD’2006». May Kyiv, 2006 [in Russian]. 6. Molchanov I.N, Khimich A.N., Popov A.V., Gerasi mo- 15. Molchanov I.N, Khimich A.N., Mova V.I., Pe re voz chi- va T.A., Nesterenko A.N., Yakovlev M.F. Ob effektivnoi kova O.L., Strjuchenko V.A., Popov A.V., Chistja ko va T.V., realizatsii vychislitel’nykh algoritmov na MIMD-kom- Yakovlev M.F., Gerasimova T.A., Zubatenko V.S., Gromov- p’juterakh. Iskusstvennyi intellect (Artificial intelli gen- skii A.V., Nesterenko A.N., Poljanko V.V., Ru dich O.V., ce). 2005. 3: 75—184 [in Russian]. Yushchenko R.A., Nikolaichuk A.A., Gorodetskii A.S., 7. Khimich A.N., Molchanov I.N., Popov A.V., Chis tja ko- Slobodjan Ja.E. Geraimovich Ju.D. Chislennoe program- va T.V., Yakovlev M.F. Parallel’nye algoritmy reshenija mnoe obespechenie intellektual’nogo MIMD-komp’jutera zadach vychislitel’noi matematiki. Kyiv: Naukova dum- Inparcom. Kyiv: Naukova dumka (Scientific thought), ka (Scientific thought), 2008 [in Russian]. 2007 [in Russian]. 8. Molchanov I.N. Problemy intellektualisatsii MIMD- 16. Kaplun A.B., Morozov E.M., Olfer’eva M.A. ANSYS v komp’juterov. Kibernetika i sistemnyi analiz (Cybernetics ru kakh inzhenera: Prakticheskoe rukovodstvo. (ANSYS and systems analysis). 1998. 1: 37—46 [in Russian]. in the hands of the engineer: A Practical Guide). Mosk- 9. Molchanov I.N. Intellektual’nye komp’jutery — sredstvo va: Editorial URSS, 2003 [in Russian]. isledovanija i reshenija nauchno — tekhnicheskikh za- 17. Gorodetsky A.S., Evzerov I.D. Komp’juternye modeli kon- dach. Kibernetika i sistemnyi analiz analiz (Cybernetics struktsii (Computer models of structures). Kyiv: FACT, and systems analysis). 2004. 1: 175—179 [in Russian]. 2007 [in Russian]. 10. Sergienko I.V., Molchanov I.N., Khimich A.N. In tel lek- 18. Khimich A.N., Poljanko V.V., Popov A.V., Rudich O.V. Re- tual’nye tekhnologii vysokoproizvoditel’nykh vychisle- shenie zadach rascheta prochnosti konstruktsii na MIMD- nii. Kibernetika i sistemnyi analiz (Cybernetics and sys- kompjutere. Iskusstvennyi intellekt (Artificial Intelli gen- tems analysis). 2010. 5: 64—176 [in Russian]. ce). 2008. 3: 750—760. 11. Molchanov I.N, Mova V.I., Strjuchenko V.A. Intel lek- 19. Bilous M.V. Finite solver Nedra-3D: modeling of water filt- tual’nye komp’jutery dlja issledovanija i reshenija nauch- ration in the space-planned staging. In Proc. The Inter na- no-tekhnicheskikh sadach — novoe napravlenie v razvi- tional Conference «The Issules of Calculation Op ti mizati- tii vychislitel’noi tekhniki. Zv’jazok (Communication). on (ISCOPT XLII)», September 21—25, 2015, Transcar- 2005. 7: 45—46 [in Russian]. pathian region, Mukachevo district, town. Chy nadiyevo. 12. Perevozchikova O.L., Tulchinsky V.G., Yushchenko P.A. 20. Velikoivanenko E.A., Milenin A.S., Popov A.V. Sido- Postroenie i optimizatsija parallel’nykh komp’juterov ruk V.A., Khimich A.N. Metody i tekhnologii paralel’- dlja obrabotki bol’shikh objemov dannykh. Kibernetika nykh vychislenii dlja matematicheskogo modelirovanija i sistemnyi analiz (Cybernetics and systems analysis). naprjazhonno-deformirovannogo sostojanija konst ruk- 2006. 4: 117—129 [in Russian]. tsii s uchetom vjazkogo razrushenija. Problemy uprav le- 13. Molchanov I.N., Perevozchikova O.L., Khimich A.N. nija i informatiki (Journal of Automation and Informa- Inparcom-16 — Intellektual’naja rabochaja stantsija. Ki- tion Sciences). 2014. 6: 42—52 [in Russian].

26 ISSN 2409-9066. Sci. innov. 2016, 12(5) Intellectual Personal Supercomputer for Solving R&D Problems

О.М. Хіміч, І.М. Молчанов, В.І. Мова, А.Н. Химич, И.Н. Молчанов, В.И. Мова, О.О. Ніколайчук, О.В. Попов, Т.В. Чистякова, А.А. Николайчук, А.В. Попов, Т.В. Чистякова, М.Ф. Яковлев, В.Г. Тульчинський, Р.А. Ющенко М.Ф. Яковлев, В.Г. Тульчинский, Р.А. Ющенко Інститут кібернетики ім. В.М. Глушкова НАН України, Институт кибернетики им. В.М. Глушкова НАН Украины, просп. Академіка Глушкова, 40, Київ-187, 03187, Україна; просп. Академика Глушкова, 40, Киев-187, 03187, Украина тел.: +38 (044) 526-60-88, факс +38 (044) 526-41-78 тел.: +38 (044) 526-60-88, факс +38 (044) 526-41-78 ІНТЕЛЕКТУАЛЬНИЙ ПЕРСОНАЛЬНИЙ ИНТЕЛЛЕКТУАЛЬНЫЙ ПЕРСОНАЛЬНЫЙ СУПЕРКОМП’ЮТЕР ДЛЯ РОЗВ’ЯЗУВАННЯ СУПЕРКОМПЬЮТЕР ДЛЯ РЕШЕНИЯ НАУКОВО-ТЕХНІЧНИХ ЗАДАЧ НАУЧНО-ТЕХНИЧЕСКИХ ЗАДАЧ Разработан новый отечественный интеллектуальный Розроблено новий вітчизняний інтелектуальний пер- персональный суперкомпьютер гибридной архитектуры сональний суперкомп’ютер гібридної архітектури Ін- Інпарком_pg, предназначенный для математического мо- парком_pg, призначений для математичного моделюван- делирования процессов в оборонной отрасли, отраслях ня процесів в оборонній галузі, галузях машинобуду- машиностроения, строительства и т. д. Создано интеллек- вання, будівництва тощо. Створено інтелектуальне про- туальное программное обеспечение для автоматического грамне за без пе чення для автоматичного дослідження та исследования и решения задач вычислительной матема- розв’язування задач обчислювальної математики з на- тики с приближенными данными различной структуры. ближеними даними різної структури. Реалізовано при- Реализовано прикладное программное обеспечение для кладне програмне забезпечення для математичного мо- математического моделирования задач в строительстве, делювання задач в будівництві, електрозварюванні та про- электросварке и процессах фильт рации. цесів фільтрації. Ключевые слова: математическое моделирование, ин- Ключові слова: математичне моделювання, інтелек- теллектуальный компьютер, гибридная архитектура, вы- туальний комп’ютер, гібридна архітектура, обчислюваль- числительная математика, приближенные данные. на математика, наближені дані. Received 23.05.16

ISSN 2409-9066. Sci. innov. 2016, 12(5) 27