Aristotle University Of Thessaloniki (AUTH)
Προσομοίωση και ανάλυση της απόδοσης μεθόδων ανάθεσης εργασιών σε υπολογιστικά πλέγματα
Abstract
dc:descriptionGrid systems integrate heterogeneous distributed resources that are not subject to centralized control, creating powerful infrastructures that can provide a variety of services. This thesis studies the allocation of jobs to resources in computational grids, with the use of simulation. The job allocation occurs at two levels. At the first level, the grid scheduler dispatches the jobs to the distributed independent sites, and at the next level, the local scheduler in each site allocates specific processors in order to start the actual process of jobs. Some of the challenges, which are taken into account in this thesis, are the effectiveness of resource management, the variability of job service demands, the system’s performance and the energy consumption. The main objective of the grid scheduler is the discovery and selection of the most appropriate sites and the efficient load sharing. To accomplish this task, the grid scheduler must have some sort of static information about the characteristics of sites. Additionally, it can utilize dynamic site load information in order to watch their state. In the present thesis, policies that utilize dynamic site load information are examined. Two different hybrid policies, which combine two policies, are proposed and evaluated. The first hybrid policy, which is applied in the case where the grid scheduler does not know in advance the service demands of jobs, selects a scheduling type based on the arrival time of a job. The second hybrid policy, which is applied when the grid scheduler is aware of the service demands of jobs, selects the scheduling policy based on a job’s demand. Jobs that are submitted to a computational grid can show high variability regarding their service demands, due to the different user demands. Schedulers must be able to deal with such a situation, especially when service demands are unknown. The performance of various policies at grid and local level, when service demands of jobs are highly variable, is studied in the present thesis. Different levels of variability are examined, and simulations are conducted in order to find the most effective combinations. The frequent transmission of dynamic load information from sites to the grid scheduler results in heavy network traffic and overhead. Two new site allocation policies are presented. One of their main advantages is that they require less frequent dynamic load information in order to operate. They are evaluated under the load balancing among the heterogeneous sites, in addition to performance. Besides performance evaluation, energy consumption is another issue that requires attention. In order to study the case where the energy consumption is taken into account, three local job scheduling polices in a cluster are proposed and evaluated. The cluster involves processors with different performance and power characteristics. One policy is optimized for energy conservation, while the other two policies are optimized for high performance. Simulation experiments are used to discover the most suitable policy according to system’s status and the desired priority (high performance or energy conservation
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zikos, Stylianos
- Ζήκος, Στυλιανός
Subjects
dc:subject × 12- Προσομοίωση διακριτών γεγονότων
- Υπολογιστικά πλέγματα
- Πολιτικές ανάθεσης πόρων
- Χρονοδρομολόγηση δυο επιπέδων
- Discrete-event simulation
- Computational grids
- Resource allocation policies
- 2-level job scheduling
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/25040