CASPIAN JOURNAL
MANAGEMENT AND HIGH TECHNOLOGIES
The simulation models of clusters and GRID-systems
 |
Read | Gaevoy Sergey V., Fomenkov Sergey A., Lukyanov Viktor S. The simulation models of clusters and GRID-systems // Caspian journal : management and high technologies. — 2015. — №1. — pp. 166-180. |
Gaevoy Sergey V. - post-graduate student, Volgograd State Technical University, 28 Lenin Av., Volgograd, 400005, Russian Federation, gaevserge@mail.ru
Fomenkov Sergey A. - D.Sc. (Engineering), Professor, Volgograd State Technical University, 28 Lenin Av., Volgograd, 400005, Russian Federation, saf@vstu.ru
Lukyanov Viktor S. - D.Sc. (Engineering), Professor, Volgograd State Technical University, 28 Lenin Av., Volgograd, 400005, Russian Federation
This paper describes the management of strategies (heuristics) for job distribution in Grid-system. Grid-system is assumed to consist of clusters. Each cluster system is represented as a queueing system (QS) that executes incoming jobs. An unlimited queue is assumed because of a requirement to execute all jobs. Strategies make allowance for balancing workload between clusters of Grid-system. In this paper the deterministic and stochastic simulation model for cluster and Grid-system are described. Both models are based on real parallel workloads (represented as SWF-logs). These parallel workloads are provided by Parallel Workloads Archive for free. In this paper each workload consists of triplets. Each triplet contains an income time, a job width and a job length. The width is an amount of faculty units being required to execute the job. The length is the time the job needs to be executed on some platform. The square (the difficulty) is a product of the length and the width. Some models for stochastic approximation of workloads are proposed. These models are accompanied by some distributions. A list of ten heuristics is given. These heuristics make allowance for queue length, queue width, system loading, system using etc. User runtime estimates are very inaccurate values. They are very different from real runtimes. Because of that we shouldn't use the heuristics containing such estimates. The deterministic and stochastic models show a similar result so that we can presume both model are correct. If a SWF-log contains some anomalies then the results can be different. That means an information from the log do not fully describe the cluster behavior. The developed models are aimed at selecting the most effective strategy of job distribution.
Key words: грид-системы, брокер заданий, имитационное моделирование, детерминированная имитационная модель, распределение заданий, эвристики распределения заданий, системы массового обслуживания (СМО), вычислительный кластер, parallel workloads, Grid-systems, resour