Translated Abstract
With the wide adoption of real-time system in many areas, and the applications become more and more complicated, so it is very necessary to build a platform with great capabilities to meet those requirements. Because cluster has many benefits such as low cost, high performance and scalability, it must have wide development prospects to make the cluster meet the real-time constraints and apply it into real-time application area.Real-time cluster is not the same as common traditional cluster, it has strict real-time constraints, and it is necessary to study system framework, networks, task scheduling etc. In this thesis, we take task scheduling that is the most important problem of real-time system as research object.In this thesis, we analyze the scheduling characteristic of real-time cluster system, and propose the two-level scheduling model according to the coarse grain of task in real-time scheduling. The scheduling problems of real-time cluster are divided into two parts in this model, the assignment on cluster level and the dispatcher on node level. Our model stands out different emphasis in different level, benefit the future research in scheduling problems of real-time cluster.In the research of the dispatcher on node level, based on the analysis of time-critical request of operational system needed by the real-time cluster, this thesis presents a Linux real-timing solution of real-time cluster. According to the characteristics of real-time process on the cluster node, we realize a new RTCS real-time scheduler based on the proportional-share CPU algorithm, give an efficient solution which meets requirements of the dispatcher on node level. In the research of the assignment on cluster level, this thesis proposes a dynamic RCDB algorithm according to requirements of dynamics, predictability and flexibility on this level. Firstly, we propose a new scheme that examines the real-time constraints of tasks before heuristic scheduling, and implement the method of the time-critical examining of the tasks with precedence constraints, reducing the response time of scheduling algorithm. Secondly, by suggesting the policy of degradation of Qos (Quality of Service) of the task, RCDB algorithm obviously improves the scheduling success ratio and system benefit obtained by schedule. Finally, RCDB algorithm can achieve the stable scheduling performance by introducing of feedback.
Corresponding authors email
