An Adaptive Scheduler for Grids

Abstract

The recent development of telecommunication infra-structures such as the world-wide networks, interconnecting millions of computers spread all over the world, has made possible the coordinated use of a large amount of heterogeneous, weakly connected computational resources. This new area, known as Grid computing, is currently the focus of several research activities and projects. However, as in every new domain of research, in this one there are many unsolved questions, in particular those related to the management of the processing load inside the system. In this work, the problem of balancing processing loads on a Grid is approached by the introduction of the Generational Scheduling with Task Replication (GSTR) algorithm. A comprehensive set of simulations and tests is carried out in order to validate the proposed solution. A methodology for calculating and analyzing speed-up and efficiency ratios on heterogeneous groups of computers is also shown.

Towards the Scheduling of Multiple Workflows on Computational Grids

“…The task scheduling has been studied in homogeneous systems [23,32,37,50]. With the development of heterogeneous systems and the Grid, it has been necessary to consider the heterogeneity in processing power and bandwidth [1,4,8,16,19,31,38,44,48].…”

Modeling and Supporting Grid Scheduling

“…Important requirements of this model are the timeframe-specificity of predictions, dynamicity, and adaptivity [5,35]. Different data models are used to describe resources, ranging from flat files to relational, hierarchical, object-oriented, and semistructured [12].…”

“…As a consequence, the model does not make assumptions on the dynamicity of the scheduling processes. For instance, dynamic scheduling processes with re-scheduling [5,18,35] can be employed where the scheduler is invoked initially and then, during the execution of jobs, it is invoked repeatedly as a consequence of the occurrence of new events. This way, it can reschedule running tasks (if mechanisms are supported) and tasks that have not been executed.…”

“…Service level agreements (SLAs) are typically considered as suitable means to support the necessary reliability. Thus, resource reservations and qualityof-service guarantees are being pointed out as a major requirement of modern GRM systems [36,38]; however, even in the absence of SLAs, some other approaches have been proposed to improve reliability and performance of resources including, e.g., performance prediction, job replication, and dynamic adaption to performance changes [35,42,50].…”

Job Allocation Strategies with User Run Time Estimates for Online Scheduling in Hierarchical Grids

“…Various scheduling systems have already been proposed and implemented in different types of Grids [1][2][3][4][5][6][7][8]. Most academic studies either propose a completely distributed resource management system, see, for instance, [9], or suggest a central scheduler, see [10,11], while real installations favour a combination of decentralized and centralized structures, see [12].…”