Home         Authors   Papers   Year of conference   Themes   Organizations        To MES conference

Dynamic management of computations in distributed systems  

 Lupin S.A.
 Pachin A.
 Kostrova O.
 Fediashin D.
Date of publication

 In this article we solve a problem of increasing the ef-ficiency of the distributed computing environment by using a method that dynamically redefines parameters of the subtasks which are running on the network nodes. The implementation of this method is based on multiple run of the optimization applications. The results of math-ematical modeling and practical implementation show practical effectiveness of this method.
There is a task of using effectively the resources a GRID system which exploits one single integrated envi-ronment. In order to eliminate some problems of BOINC systems for small projects, the dynamic computing control method, which increases the efficiency of the nodes in the GRID systems, can be used.
The idea of the proposed method is to launch distributed applications repeatedly when redefining the tasks each time. This approach can be used for solving function optimization problems when the variants enumeration algorithm is applied.
The algorithm can be described as follows: first, we define the searching area and distribute it uniformly among the GRID nodes, and then we choose a sub-area, in which we get the best result. Next we decrease the step size of function parameters changing and use the afore-mentioned are to continue optimum searching. The pro-cess continues until the required accuracy of solution or time limit is reached.
In order to evaluate the effectiveness of the proposed method we built a MATLAB mathematical model that helps us to find the optimal parameters. The results of the computation experiments show the possibility of using this method for solving optimization problems.
To find the extremum of the function we use the most universal search algorithm. Despite the fact that the execution time is 11 hours on one node, we use this algo-rithm because of the following advantage. It is invariant to the test function; therefore it can be used when the function is not differentiable on the entire domain which is mandatory for the gradient methods that can solve such problems more quickly.
Experimental results show the advantages of using the proposed algorithm. In order to localize the search area at first stage we use multiplied by 10 the initial step size. This allows us to reduce the execution time by 100 times and to select a perspective search area. Note that we limit this area by using only one variable; it is possible to limit the area by using several ones when a large number of nodes are used. On the last stage we apply the algorithm to a previously limited area. Since the runtime is proportional to the number of analyzed points, the distinguished acceleration is equal to 25, which is consistent with the theoretical evaluation.
 distributed computations, GRID, mathematical modeling
Library reference
 Lupin S.A., Pachin A., Kostrova O., Fediashin D. Dynamic management of computations in distributed systems // Problems of Perspective Micro- and Nanoelectronic Systems Development - 2016. Proceedings / edited by A. Stempkovsky, Moscow, IPPM RAS, 2016. Part 2. P. 185-189.
URL of paper

Copyright © 2009-2019 IPPM RAS. All Rights Reserved.

Design of site: IPPM RAS