Translated Abstract
Rendering application has the hybrid characteristics of computational intensive and data intensive. With the rapid development of high performance computing (HPC) and digital multimedia, render farm equipped with high performance cluster utilizes the application characteristic of parallelism among rendering frames, and it becomes an important mean to solve the problem of time consuming for rendering process, Internet technology and cloud computing technology improve the extension ability of render farm. scale expansion for render farm ensures the high processing ability, but if there is no optimization strategy for rendering system, the rendering system cannot give full play to the computing power. How to optimize job scheduling strategy for render farm, assign the rendering tasks evenly to rendering nodes and improve the resource utilization and performance of rendering system is the important challenges for the research on performance optimization of render farm. In addition, the expansion of render farm also means the increase of system energy consumption, how to design the reasonable scheduling strategy to reduce the energy cost is also a very challenging problem.
To overcome the above challenges, extensive and intensive research is carried out on optimizing job scheudling and energy consumption of rendering applications in a render farm. This study is carried out from three perspectives: rendering job scheduling stragety, energy optimization oriented task scheduling stragety and scheduling stragety in hybrid render farm. The content of this dissertation is as follows.
1) Aiming at the problems of job starvation and resource fragmentations, speedup bottleneck with multi-thread rendering, load imbalance with equipartition by frames method in traditional job scheduling strategy for rendering farm, a hierarchical rendering job scheduling strategy for render farm is proposed. The strategy divides the job scheduling in render farm into two levels, the method of rendering job scheduling level considers the feature of non-dependencies among frames for render application and establishes the rendering job scheduling model, and the rendering jobs are scheduled by priority. When the rendering resources are limited, the job with low priority will be scheduled in advance by the way of modifying the priority on the condition that it cannot delay the job with high priority, and the job with high priority and high resources demand will reserve rendering resources. If the rendering resources are enough for the job with high priority, the job with low priority will be paused and the job with high priority will be rescheduled. The method of rendering task distribution level adopts the “rendering unit” as the basic unit for resources partition granularity, and the scene geometry based method for choosing key frames is proposed. According to feedback of the resource usage information by pre-rendering the key frames and the features of correlation among frames, the frames sequence has been distributed evenly to the rendering units combined with the intra-frame partition method. Experimental result shows that when compared with existing methods, the proposed strategy can improve efficiently the resource utilization rate, balance the workload among rendering nodes effectually and reduce the job completion time and response time effectively, and also ensure the fairness.
2) Aiming at the issue that the idle of render nodes may causes a great waste of energy consumption in the job scheduling strategy for cloud rendering system, the energy optimization oriented rendering tasks scheduling strategy is proposed. The strategy considers the task running energy consumption and the energy consumption from idle nodes comprehensively and combined the feature of independence among frames. The rendering task energy consumption model is presented. Based on the model, the tasks scheduling queue is split to sub queue. With the simulated annealing ideology, the strategy optimizes the scheduling of the subsequence tasks to improve the utilization ratio of the rendering nodes and reduce the idle time of rendering nodes, and then achieves the reduction of energy consumption for the overall system. The experimental results show that, compared with the existing FCFS strategy and EMRSA, the proposed strategy has improved the performance of energy optimization effectively in the test for real rendering scene and possesses better expansibility of tasks and node scales.
3) Aiming at the issue that the render farm with local resources may delay the job completion time caused by the insufficient computing ability, the budget oriented rendering job scheduling strategy in hybrid render farm with hiring the cloud resources is proposed. Under the constraint with Deadline and budget, the strategy set the budget factor to control the number of cloud resources hired by users, which ensures that the job can finished on time and provides an economic way to hire the cloud resources. In addition, Aiming at the issue that the rendering users cannot verify the reliability of rendering source data and result images and the security ensured by the cloud providers, the active verification trust method for cloud resources is proposed. With the purpose of trust verification, the method verify reciprocally the “spy frames” by multiple cloud rendering farms. Experimental result shows that when compared with existing Cost-Opt method, the proposed approach can satisfy the demand of Deadline and also obtain the high cost efficiency. In the meantime, the proposed method for reliability verification could verify availably the malicious action of prolonging rendering time from cloud render farm.
In this dissertation, the performance and energy consumption optimization issues in render farm are studied by combing theoretical analysis and engineering practice. Three approaches are proposed: rendering job scheduling strategy, energy optimization oriented task scheduling strategy, job scheduling strategy for hybrid render farm and verifing method for trust of cloud rendering, respectively. Such approaches have important theoretical value for performance optimization and energy management of render farm systems.
Corresponding authors email
