Aristotle University Of Thessaloniki (AUTH)
Αποτίμηση της απόδοσης πολιτικών χρονοδρομολόγησης παράλληλων εργασιών σε συστήματα πλέγματος με μοντέλα προσομοίωσης
Abstract
dc:descriptionThe demand for computing resources of modern applications has been increased significantly as compared to the past. Meeting the modern demands with existing resources has become feasible by organizing the available resources appropriately and sharing them. A grid system coordinates resources that are not subject to centralized control using general-purpose protocols to deliver non-trivial qualities of service. The development of efficient scheduling algorithms is required in order to efficiently utilize grid systems. The main subject of this thesis is the performance evaluation of parallel job scheduling policies in grid systems by means of simulation models. The workload characteristics are responsible for the performance of the system. Therefore, the system must implement the necessary algorithms and techniques that will provide adequate flexibility to cope with all possible situations. For this reason we examine the performance in the case of workload which comprises Bags of Gangs. We also utilize a migration schema to balance the load in the heterogeneous system. Fragmentation is a situation which frequently occurs in the case of gang scheduling in systems. We suggest a migration technique which alleviates this problem and improves the system performance drastically. We examine this technique under various workloads. We consider the presence of critical sporadic jobs and we also examine the impact of the variability in gangs’ service time on the scheduling performance. Load balancing plays a significant role on the overall system performance. We examine a new grid scheduler which implements periodic feedback. The goal of this scheduler is to evenly distribute the load across the available clusters. The scheduler is considered suitable for gang scheduling. This thesis also analyses a workload model which comprises gangs with precedence constraints. In this case a subsequent gang can be dynamically generated in the system after the execution of the initial gang based on affinity information which resides in the caches of the previously seized processors. Different gang scheduling policies are examined under the new workload model. Another interesting topic which we consider in this study is the scheduling of gangs in multi-core clusters. We modify the gang scheduling policies and make them suitable for homogeneous and heterogeneous multi-core clusters. Furthermore, we provide a migration schema which is suitable for scheduling gangs in multi-core clusters. We investigate the benefits and differences between the performance of single-core algorithms and their multi-core counterparts. Finally, we deal with the case of heterogeneous multi-core clusters and present the impact on the performance of our policies. To sum up, the main contribution of this thesis is the development of new scheduling algorithms for parallel jobs in order to improve the performance of grid systems. The suggested scheduling algorithms allow for better resource utilization and have a significant impact on grid systems’ performance. We achieve this by implementing various techniques such as improved grid scheduling algorithms with periodic feedback and new migration schemas. We also consider and examine complicated workload models which include Bags of Gangs, gangs with precedence constraints and critical jobs. Finally, we suggest new gang scheduling and load balancing algorithms for multi-core clusters.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papazachos, Zafeirios
- Παπαζάχος, Ζαφείριος
Subjects
dc:subject × 20- Υπολογιστικό πλέγμα
- Χρονοδρομολόγηση
- Απόδοση
- Μοντέλα προσομοίωσης
- Παράλληλες εργασίες
- Εξισορρόπηση φόρτου
- Μοντελοποίηση
- Κατανεμημένα συστήματα
- Grid computing
- Scheduling
- Performance
- Simulation models
- Parallel jobs
- Load balancing
- Modeling
- Distributed systems
- 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/25041