A Hierarchical Scheduling Policy for Large-Scale Rendering

With the development of high performance computing, especially after the emergence of petaflop system, people pay more attentions on how to use these resources in an effective way and what application could be deployed on them except for scientific computing. Given the properties of high performance computing, rendering is one kind of application which is very suitable for distributed computing due to intensive computation and massive data access. A rendering task could be decomposed into multiple independent subtasks and submitted for parallel execution. But because of the heterogeneity and unbalanced load in the computing environment, there often appear such circumstances where there are severe load unbalances. Moreover, the high performance platform is not exclusively used for rendering on account of the sharing of computing resources. For rendering there are some new available resources at a given time and at another point those resources will possibly become unavailable likewise. In this paper, we propose a hierarchical scheduling policy for large-scale rendering to resolve above problems. And afterwards we focus on load balance layer which is the key part in our policy. In order to make use of the resources in a more effective style, the load balancing for rendering based on feedback of the runtime information is proposed. We evaluate the proposed approach comparing with non-strategy approach. The results show our approach can be competitive choice for rendering schedule and outperforms non-strategy by 10.2%-23% in terms of completion time.