
Registration


Opening of the GraphHPC2017
V. Voevodin, Corresponding member of the Russian Academy of Sciences, MSU RCC
A. Simonov, JSC NICEVT


A. Pozdneev, IBM
Hide
Graph databases are increasingly popular in managing the information where the relationships between the data entities are of highest priority. However, a technical task of deploying, managing, and maintaining a graph database onapremise is decoupled from the process of solving an applied problem. IBM Graph is a graph databaseasaservice available on the IBM Bluemix cloud platformasaservice. IBM Graph is built upon opensource components while featuring highavailability and scalability ondemand. In this talk, we will introduce the main concepts behind IBM Graph and show how to leverage its API and the Bluemix console GUI.


G. Fedorov, Intel Software
Hide
Intel’s goals for it is to provide the technology for BigData processing and Machine Learning challenges available for broader range of researchers and developers, and their results available for consumers , and for domain experts without specific skills in programming and computational sciences. Another priority is to increase efficiency of implementations of machine learning algorithms when running on Intel processors. Intel DAAL library solves this problem.
The presentation will show the key capabilities and features of the use of the library, is an example of Intel DAAL use to solve a "classical" problem of recognition of handwriting and some comparative results of performance and scalability.


The virtual cluster with low latency network on a stack of technologies: qemu / KVM, Сeph, OpenStack
A. Nikolaev, HPC Hub
Hide
Simulation of complex physical processes is an important technological option for various companies in the fields of BigData, oil and gas services, pharmaceutical, automotive, aircraft and other types of engineering.
The creation of the cluster systems where the compute node is an industrial general purpose server with its own operating system from the GNU / Linux family, is a widely used approach now for building computation systems capable to solve complicated numerical problems
Because of the relatively small amount of clustered systems, there is no specialized cluster OS and GNU / Linux in a clustered version of the OS is highly specific, requiring highly skilled staff.
Virtualization layer allows users to create "virtual cluster" within minutes and makes it possible to provide businesses and research groups with a “cluster on demand” service. These virtual clusters within the same OpenStack infrastructure are absolutely independent.
User programs within those clusters may vary as it’s needed to the user without additional arrangements with neighbouring users and infrastructure owners, and logic devices,containing user data are not reachable for other virtual clusters.
In most cases, virtualization results a minimal loss of compute power (<1%), for example, tNavigator comparative tests by RFD and ProMAX from LandMark did not show any performance loss in our tests.
However, specialized tests (OSU benchmark) show low latency network overhead of virtualization is not more than 20% for the synchronization operations.


Coffee Break


Learning PageRank for Web Page Ranking
M. Zhukovskii, L. Bogolubsky, P. Dvurechensky, A. Gasnikov, G. Gusev, Y. Nesterov, A. Raigorodskii, A. Tikhonov, MIPT / Yandex
Hide
The most acknowledged methods of measuring importance of nodes in graphs are based on random walk models. In particular, PageRank is applied for the problem of ranking web pages. Despite undeniable advantages of PageRank and its modifications, these algorithms miss important aspects of the graph that are not described by its structure. In contrast, a number of approaches allows to account for different properties of nodes and edges between them by encoding them in restart and transition probabilities. In the talk, we present novel optimization methods of learning parameters of such models.


Mathematical and Computer Science problems in Bioinformatics
V. Lyubetsky, Institute for Information Transmission Problems, Russian Academy of Sciences / Department of Mechanics and Mathematics, Lomonosov Moscow State University
Hide
The report is addressed to the following problems.
1) Given n natural numbers with repetitions, determine if a whole set of them can be split into two parts with equal sum of the numbers in each part.
2) Global and local alignments of given multiple sequences.
3) Given are graphs a and b and a fixed list of natural operations on graphs. Find the shortest sequence of the operations converting a into b.
4) How to define and compute the distance between graphs.
34a. Big integer linear programming (ILP) tasks related to these problems.
5) What is an average graph for the given set of graphs.
6) Describe mode switchings in the functioning of dynamical systems of some types important in biology.
6a. Competition of opposite flows.
6b. Combination of three and onedimensional diffusions.
6c. The pursuit problem in biology.
In all cases above, a lowdegree polynomial (practically, nearlinear) algorithm to be found to solve the problem.
The important role in these studies play programming for supercomputers (especially distributed memory systems) and Big Data processing. Memory organization and distribution of the interim tasks between computational processes is also important. In particular, dimensionality reduction of corresponding ILP problem. Since the problem often cannot be solved by exact algorithm, a simulation modeling and heuristic algorithms can be of interest as well as methods of comparison of different heuristic algorithms for the same problem.


Graph processing: main concepts and systems
M. Chernoskutov, Krasovskii Institute of Mathematics and Mechanics
Ural Federal University
Hide
Nowadays, supercomputing research community are aware of vast amount of big graphs processing systems. Systems like Parrallel BGL, Pregel, Giraph are widespread now. All of these systems are differs a lot on the (typical) problem size, level of parallelization, set of algorithms and other characteristics. Thus, there is important question of searching an “ideal” graph processing system, which addresses all required user tasks in most efficient way.


V. Bakhtin, A. Kolganov, V. Krukov, N. Podderyugina, M. Pritula, O. Savitskaya, Keldysh Institute of Applied Mathematics Russian Academy of Sciences / Lomonosov Moscow State University
Hide
DVMsystem was designed to create parallel programs of scientifictechnical calculations in CDVMH and FortranDVMH languages. These languages use the same model of parallel programming (DVMHmodel) and are the extensions of standard C and Fortran languages with parallelism specifications, implemented as compiler directives. DVMHmodel allows to create efficient parallel programs for heterogeneous computational clusters, which nodes use as computing devices not only universal multicore processors but also can use attached accelerators (GPUs or Intel Xeon Phi coprocessors). This report discusses new means to work with irregular grids, which were implemented in the CDVMH compiler recently. Using the developed extension can significantly ease irregular grid applications parallelization on cluster.


Lunch Break


Parallel algorithms for sparse matrix ordering
A. Pirova, Lobachevsky State University of Nizhni Novgorod
Hide
This report considers the problem of finding an ordering of sparse matrix that minimizes the Cholesky factor fillin. For this purpose, heuristic approaches based on graph algorithms are applied. There are several libraries that are widely used in applications for sparse matrix ordering, i.e. ParMETIS and PTScotch. These libraries are designed for distributed memory systems, take advantage of MPI and demonstrate reasonable speedup. In 2015 mtmetis, a sharedmemory version of Metis library, was introduced. At the same time we developed an opensource library PMORSy that contains parallel multilevel nested dissection algorithm for sharedmemory systems. Parallel processing is done in a taskbased fashion employing threading techniques on shared memory using OpenMP 3.0 technology. Current research is focused on developing an algorithm best suited for distributedmemory systems. The algorithm is similar to that of PTScotch. The experimental results on the symmetric matrices from the University of Florida Sparse Matrix Collection will be presented. They prove that PMORSy is competitive with the ParMetis and PTScotch libraries in factor fillin and performance.


Graph theory in memorybound problems
M. Ermakov, A.Ishlinsky Institute for Problems in Mechanics RAS
Hide
Two practical problems of graph theory which are required to be solved on a cluster for a limited memory are given.


Seam line optimization for remote sensing image mosaicking based on graph shortest path finding
A. Sudorgin, JSC “Research Institute of Precision Instruments” (JSC “RI PI”)
Hide
Seamline determination is important step of remote sensing image mosaicking. Automatic seam line selection problem using graph shortest path finding is discussed in this paper. Approaches to graph construction from image similarity cost function, graph memory layout and parallel seam line optimization using shortest path on graph are presented.


On the paradigm of a universal parallel programing language by example of SSSP and BFS problems
A. Klimov, Institute of design problems in microelectronics of Russian academy of sciences
Hide
There are many different platforms for parallel computing, and each requires writing a program almost from scratch. Is it possible to produce all of these different implementations from a unified description of the computation in a "universal language"? The report will discuss the draft of such a language. It is based on the dataflow semantics. First, we write in it an abstract scheme. Then, bringing additional information about the process of computing, we can get different specific programs for different platforms. The language and its possibilities are shown in the report by an example of problems SSSP (Single Source Shortest Path) and BFS (BreadthFirst Search).


Early experience with integrating Charm++ support to GreenMarl domainspecific language
A. Frolov, DISLab / JSC NICEVT
Hide
In the talk there will be presented the most valuable results of the research on using of a parallel programming framework Charm++ for largescale graph applications. An objectbased messagedriven computational model of Charm++ allows to effectively implement graph algorithms in a vertexcentric approach while preserving high scalability and performance. For enabling a further productivity improvement a compiler for domainspecific language GreenMarl for static graph analysis has been ported to Charm++, which also will be presented in the talk.


GraphHPC2017 contest closing


GraphHPC2017 closing
