Research On Resource Management And Scheduling Strategy Based On Energy And Cost

Large-scale computing system has improved its performance through parallel processing and high-speed network, which is extensively applied not only in the fields of petroleum exploration, national defense, space navigation, weather forecast, military simulation and industrial design, but also in the emerging areas of finance, telecommunication, government, education and online game, etc. It is the most powerful weapon of exploring the unknown.With the increasing scale of systems, high-end computing starts to pursue high productivity instead of just high performance. Energy management not only is the research hotspot of battery-driven devices, but also becomes the critical issue for data centers and supercomputing centers due to their high operating cost, low reliability and environmental effects. With the technical innovation, the acquisition cost of servers is trending down, while energy related cost including capital charges, operating expenses and environmental consequence is trending up. Therefore, improving power efficiency and achieving green computing are the main technology development trends of computer and IT industry.Energy index as system demand is affected by execution efficiency of applications on the one hand, while on the other hand it will also influence users in the form of resource or service usage cost. One way of indicating usage cost is the price. Currently the existing pricing schemes lack the accurate mathematical model and are rarely related with energy consumption. In this way, energy conservation of systems cannot be maintained and resource or service usage cost of users cannot be optimized, which hurts the interests of both parties. Thus, the research on pricing based cost management is necessary. In addition, from the global perspective, the load unbalancing of systems results in the cost increase and efficiency reduction, while the resource or service usage cost of users has also not dropped but increased owning to the conflicting use of the high-quality resources and the lack of cooperation. Hence, investigating load balancing based on cooperative mechanism is an efficient way to reduce the cost of systems and users.System productivity depends on its hardware as well as the resource management system deployed on the architecture. Resource management is in charge of responding the requests of users, effectively scheduling tasks and reasonably assigning resources. Task scheduling as the core of resource management is responsible for task ranking and resource allocation on a set of processing elements with arbitrary characteristics. Multi-task scheduling problem is not well solved due to its NP hardness, the diversity of environments, new demands of applications and compromise of scheduling objectives, especially in the context that IT energy consumption becomes an important factor of affecting global energy and environmental change.The dissertation, supported by National863Program and Natural Science Foundation of China, aims at the problem of energy management and cost management in large-scale computing system with the demand of high productivity. Using the way of resource management and task scheduling, it abstracts several challenging scheduling problem through the analysis on power-aware resource management approaches, task scheduling models and the limitations of current scheduling algorithms. The solution of these scheduling problems can effectively fill in the gap and solve the deficiency of existing work, which takes into account the variations of different execution environments, the limitations of different scheduling techniques, the computation-intensive or data-intensive property of different applications and the conflicts of different QoS constraints and performance goals, etc. With the increasing trend of IT energy consumption, the widespread use of cloud-computing based business model and the continuous expansion of system scale, the studies on resource management and task scheduling which mainly focus on the issues of energy conservation, market model and scheduling cooperation have both theoretical and applicable values.The main contributions and innovations of the dissertation are as follows:Firstly, we focus on the energy-aware scheduling of independent tasks. With the rapid development of technology and expansion update of equipments, heterogeneous computing systems are more common than homogeneous computing systems. Heterogeneous system takes the full advantage of parallel processing, but also enhances the complexity and diversity of application execution. Based on the analysis on the existing scheduling algorithms and power-aware resource management approaches, we propose a dynamic power management based energy-aware deadline scheduling algorithm of independent tasks in heterogeneous systems. The algorithm is designed for long-running independent tasks in high-performance computing field, which are different from the periodic tasks in real-time systems and also in contrast to the workloads made of small individual processing units such as http requests. By analyzing the scheduling model in heterogeneous multiprocessor system and considering the deadline constraint of applications, we first prove the NP-hardness of the problem, formulate it as integer linear programming model. We also design an efficient deadline scheduling algorithm based on dynamic power management technique. Its worst-case performance is theoretically analyzed and experiments also demonstrate its effectiveness. The algorithm makes up the one-sideness of the solution based on dynamic voltage and frequency scaling method. The extensive experiments demonstrate that the proposed algorithm achieves near-optimal energy efficiency,2%-20%better in loose deadline and5%-44%better in tight deadline than EDD (Earliest Due Date)-based algorithm. In order to facilitate the marketization application of our algorithm, based on the proposed unitcost metric, we present a pricing scheme which relates cost with energy consumption, provides a way of adjusting cost for users and improves the motivation of energy conservation from the users' side, not just from the provider's side.Secondly, we study the energy-aware scheduling of dependent tasks. Since supporting the highly data-intensive workloads is becoming the key technique of the next-generation supercomputing center and data center, the specific scheduling framework and solution especially propitious to the characteristics of this field are exigent to be proposed. In order to overcome the limitations of existing work in scheduling technique, communication energy optimization, system heterogeneity and static energy consumption, various energy-aware scheduling algorithms of data-intensive applications in different systems are discussed. First we focus on homogeneous systems, construct system model, application model and power model, and propose the efficient scheduling framework. Then, based on the framework, we extend to heterogeneous systems and propose energy-aware scheduling solution considering the heterogeneity of both computation and communication resources to further trade off the relationship among precedence constraints of application, system heterogeneity and conflicting scheduling indicators. For homogeneous systems, we give different system models in terms of whether supporting dynamic voltage and frequency scaling or not and discuss their scheduling approaches respectively. Since data-intensive applications have massive data transmission, their complex precedence-constraints, data transfer time and communication energy consumption can not be ignored. We introduce the methods of reducing communication cost, such as task duplication, task clustering, dynamic mapping parameter, to satisfy the specific demands of this field. Along with the development of chip miniaturization and multi-core technology, static power increases exponentially as the number of the electric components in unit size rises. We thus introduce coarse-grain or fine-grain dynamic power management technique to reduce the non-ignorable static energy. In addition, since the performance goal and energy index conflict each other, we use reasonable conditions to make a compromise during the usage of various techniques. The experiments show that the proposed scheduling frameworks and solutions can not only adapt to the characteristics of systems and applications, but also efficiently ensure the achievement of scheduling objectives.Thirdly, pricing based cost optimization scheduling is studied. Energy index of systems influence users in the way of resource or service usage cost. In order to guarantee the interests of system and users, we propose the market-driven enactment engine scheduling framework and design a cost optimization mapping strategy based on marginal pricing and cost-gradient metric. The strategy gives a pricing scheme from the user's point of view which has important impacts on resource allocation, providers'profits and cost optimization. The pricing method uses the marginal principle in economics, gives an accurate mathematical model, considers the limited use of multiple resources and achieves the high income and resource utilization rate for the resource nodes. Cost optimization scheduling algorithm overcomes the disadvantages of graph partition method by using dynamic full-graph scheduling, and has good optimization ability. Cost-gradient metric helps quickly find the optimal or near-optimal services when the algorithm gets into a dilemma through searching the service with the maximum time reduction and minimum cost increase. An efficient pruning strategy is used to decrease the planning time of scheduling. In addition, during analyzing the mapping problem, we use a novel method called tree expansion analysis of cost matrix, which makes our analysis intuitionistic and efficient.Finally, double-layer load balancing scheduling strategy based on the cooperation mechanism is presented. The strategy considers the load balancing of both enactment engine and bottom resources, which provides the strong support for improving system performance and reduce cost. The application submitted by users is parsed by enactment engine. The tasks after parsing are assigned to bottom scheduler and completed by concrete resources. In this case, enactment engine is in fact the upper service for the application compared with the large-scale computing platform. Our strategy tries to solve the problem of many scheduling entries, lack of cooperation between multiple execution engines and conflicting scheduling of bottom resources in large-scale computing systems. On the basis of analyzing the function and load consumption of scheduling engine in detail, we introduce the real-time load models of execution engine and resource node, and the predicted load model of arrival applications. We also propose the cooperative architecture, and the gather and evaluation of runtime status about engines and resources are completed by monitor, planner and analyzer, which provide the accurate data for the computation of utility functions. Last the proposed double-layer scheduling algorithm based on performance model considers not only load balancing of scheduling engines and bottom resources, but also the QoS requirements of applications such as the execution time and communication time.Based on the above work, the optimization of static energy, comprehensive management of various resources and experimental setup are the future work.