Celková přidělená částka z MŠMT na specifický vysokoškolský výzkum na VŠB-TUO - **52 908 039 Kč**

Z toho 2.5% - 1 320 739 Kč - úhrada způsobilých nákladů spojených s organizací SGS

fakulta | přidělená částka v Kč |
---|---|

FBI | 1 172 500 |

EKF | 4 962 700 |

FAST | 3 070 000 |

FS | 8 256 000 |

FEI | 12 282 100 |

HGF | 5 433 000 |

FMMI | 6 188 000 |

VC | 10 223 000 |

CELKEM |
51 587 300 |

Kód | SP2015/105 |
---|---|

Název projektu | DPDM - Database of Performance and Dependability Models |

Řešitel | Nowaková Jana Ing., Ph.D. |

Školitel projektu | prof. Dr. Ing. Miroslav Pokorný<br /> |

Období řešení projektu | 01.01.2015 - 31.12.2015 |

Předmět výzkumu | 1. Anotace
Most complex systems in all technological areas need to be quantitatively evaluated at design phases, or when operating, for maintenance goals, for future developments, etc. These quantitative evaluations are globally done from mainly two distinct viewpoints: performance and dependability. Typical evaluations correspond to very costly numerical procedures. The reason is that most of them suffer from combinatorial explosions making hard to analyze even medium size models. As a consequence, another important research branch look for good approximations, bounding schemes, etc., with a cost much less sensitive to the size of the models. In this context, an extremely useful tool will be a database containing different models together with their exact solutions. This is of high value if one wants to develop a new approximating scheme, or even a new specialized Monte Carlo procedure (for instance, in a rare event situation), since it allows having the exact solution of something usually very hard to evaluate with standard hardware. Our goal is to build a database of models in both the performance and the dependability evaluation areas, each with the exact value of the important metrics, and to make it accessible to the research community. 2. Složení týmu a jeho kvalita Sebastian Basterrech, Samir Mohamed, Gerardo Rubino and Mostafa Soliman, Levenberg - Marquardt Training Algorithms for Random Neural Networks, Computer Journal, Oxford Journals, Vol. 54, Number 1, pp 125-135, 2011. Sebastián Basterrech, Jan Janoušek, Vaclav Snášel, "A Study of Random Neural Network Performance for Supervised Learning Tasks in CUDA," Intelligent Data analysis and its Applications, Volume II, Advances in Intelligent Systems and Computing Volume 298, 2014, pp 459-468. (Best paper award of the First Euro-China Conference on Intelligent Data Analysis). Janoušek, Jan.: Classification via Nearest Prototype Classifier Utilizing Artificial Bee Colony on CUDA. Advances in Intelligent Systems and Computing Volume 299, 2014, pp 21-30 (2014) Platos, J., Snasel, V., Jezowicz, T., Kromer, P., & Abraham, A. (2012, October). A PSO-based document classification algorithm accelerated by the CUDA platform. In Systems, Man, and Cybernetics (SMC), 2012 IEEE International Conference on (pp. 1936-1941). IEEE. 3. Cíle projektu a očekávané výstupy (publikace, sw, grantová přihláška, ...) A quantitative evaluation of a complex system is realised according two points of view: its performance and its dependability. Performance means that we assume the system perfect, and we look at different aspects related to the work it does (or the service it provides). Dependability ignores the work done by the system, and focus the analysis on the fact that all real systems are subject to failures and possibly repairs. In some cases, both types of aspects are considered simultaneously, and we say that a performability evaluation or assessment is done. Examples of performance metrics are response times or delays, throughputs, power, etc. Examples of dependability ones are Mean Time To Failure, reliability, availability (in many forms), maintainability, etc. The objective of this project is to build a first version of a database of models in the performance and dependability evaluation areas together with their exact numerical evaluations. The models will be large ones, “large” meaning that their evaluations are very hard or impossible to do in reasonable time with standard hardware. For instance, in both areas, many models are continuous time Markov chains living in very large state spaces. Several basic metrics correspond to solving a linear system, others a linear differential system, whose sizes are those of the corresponding state spaces, for models having millions of states (or billions, the combinatorial explosion of the state spaces is huge in these fields). The goal is to push the sizes of the models as high as possible thanks to the hardware resources of ANSELM National Supercomputing Center, while using different numerical procedures for each model in order to guarantee that solutions are correct (cross validation). For each model, we will allow some central parameter (an arrival rate, a failure rate, etc.) to take different values (then, leading to several models with the same structure and different values of their parameters), because this will be a great help to researchers in the before mentioned areas. Sequences of models where some parameter changes, increasing the size of the model, will also be built, in the same spirit. Both ideas are very useful when analyzing the behavior of approximating techniques (including Monte Carlo ones). This project will start with Markov chains in both fields, and specifically we will build different Markov models appearing in the literature and explore the limits of the CUDA system to provide accurate evaluations of important metrics. If time allows, we will also start exploring combinatorial (static) models in the network reliability area. As an indication of expected outputs, all Ph.D. students will be required to have at least two peer-reviewed articles published or accepted for publication as the first author, prior to the end of the project. Besides, as global output of the project at least two peer-reviewed journal articles must be published or accepted, prior to the end of the project. |

Členové řešitelského týmu | Sebastian Basterrech, PhD. Ing. Petr Berek Ing. Clare Schenk Ing. Jan Janoušek Ing. Tomáš Ježowicz Ing. Jana Nowaková, Ph.D. Ing. Jana Nowaková, Ph.D. prof. Dr. Ing. Miroslav Pokorný Ing. Lukáš Zaorálek |

Specifikace výstupů projektu (cíl projektu) | 4. Impakt navrhovaného projektu (aktuálnost tématu, vědecká závažnost a odborná úroveň)
According to some estimation evaluation, the fraction of papers dealing with performance and dependability evaluation methodologies represents more than 2/3 of the total number of publications in applied mathematics. In any case, the area is really huge, and the impact of such a database will be high in the research communities. It is important to underline that there are several communities: a large one dealing with static models in dependability, a huge one working with queuing systems (in performance evaluation), and leading in some cases to Markovian problems, etc. Moreover, there are many software tools available in these fields [1], [2], [3], [4], plus many proprietary ones. Having sets of models hard to analyze with standard hardware for which the exact values (of the main corresponding metrics) are available, will be very interesting for their evaluation and their comparison. This type of tool is available in other scientific areas, but not in the network reliability area, in spite of their practical importance. In France, there is something in these lines at the ISDF [5], but the models are quite small since they have been analyzed using (old) standard hardware, and they cover a pretty limited range of applications. Also, the size of the database is modest. What would be very useful is to know the exact solution of much larger models in the same families and in many other ones. That is precisely our objective. Therefore, we expect a strong impact from our contribution to the community. Several times colleagues have requested for something of that kind, thus it will be extremely useful in the associated research communities. 5. Postup řešení (dosažení cílů) s časovým harmonogram We starts solving benchmark problems using the direct method GTH specialised designed for Markov problems. Once we reached results, we starts the processing of iterative techniques. We will use the floating point accelerator equipped with Intel Xeon Phi 5110P. The implementation will be realized in the Compute Unified Device Architecture (CUDA) platform and the framework Open Computing Language (OpenCL). Additionally, over the compute nodes without accelerator, we will use the Unified Parallel C (UPC) in order to improve our computing performance on large-scale parallel machines. In Markov problems, the program structure is quite simple, essentially consisting in (row) vector matrices computations, and variations on that. An important remark is that this holds also for the differential linear systems typically appearing in dependability assessment (thanks to the so-called Uniformization or Jensen procedure). To solve the linear systems and differential linear systems present in our problems, we will use the CUDA Math Library and the linear algebra libraries available in the community. REFERENCES: [1] Dependability Institute, France, http://www.imdr.fr/en/sommaire/index.php [2] SPNM (Stochastic Petri Nets Package), http://people.ee.duke.edu/~kst/software_packages.html [3] Möbius, https://www.mobius.illinois.edu/ [4] PRISM, http://www.prismmodelchecker.org/ [5] http://queueing-systems.ens-lyon.fr/ for a recent specialized tool for solving specific queues. |

Návrh | Skutečnost | |
---|---|---|

1. Osobní náklady Z toho | 26800,- | 26800,- |

1.1. Mzdy (včetně pohyblivých složek) | 20000,- | 20000,- |

1.2. Odvody pojistného na veřejné zdravotně pojištění a pojistného na sociální zabezpečení a příspěvku na státní politiku zaměstnanosti | 6800,- | 6800,- |

2. Stipendia | 192000,- | 192000,- |

3. Materiálové náklady | 6000,- | 0,- |

4. Drobný hmotný a nehmotný majetek | 4000,- | 1010,- |

5. Služby | 30000,- | 0,- |

6. Cestovní náhrady | 89500,- | 128490,- |

7. Doplňkové (režijní) náklady max. do výše 10% poskytnuté podpory | 38700,- | 38700,- |

8. Konference pořádané VŠB-TUO k prezentaci výsledků studentského grantu (max. do výše 10% poskytnuté podpory) | 0,- | 0,- |

9. Pořízení investic | 0,- | 0,- |

Plánované náklady | 387000,- | |

Uznané náklady | 387000,- | |

Celkem běžné finanční prostředky | 387000,- | 387000,- |