A Workload-aware Load Balancing Algorithm for Cluster Rendering Platform
Qian Li,Weiguo Wu,Dun Yuan,Kang Liu,Lei Jia,Jianhang Huang
DOI: https://doi.org/10.1109/cse50738.2020.00015
2020-01-01
Abstract:Load balancing is essential for a render farm on a supercomputer platform, according to the architecture characteristics of Sunway Blue Light and rendering applications, the design of a rendering cell is presented to mitigate the I/O load of storage nodes. A workload-aware load balancing algorithm for a rendering cluster system is proposed to solve the problem in which the traditional interframe equipartition method, based on the assumption that multiple frames of the same scene are similar in terms of rendering time, cannot achieve optimization of the load balance and minimize the job completion time for real animation scenes. The algorithm characterizes the workload depending on the number of frames to be rendered and the user-specific computational resource demand. For heavy workloads, an interleaved task distribution method is proposed to take advantage of frame-to-frame coherence. For light workloads, benefiting from the fine-grained intraframe scheduling in the aspect of load balance, a load balancing method using flexible task granularity is introduced. By four scene models of different computational requirements, experimental results and detailed comparisons with the traditional interframe scheduling method and intraframe scheduling method show that the proposed strategy not only achieves near-optimal load balancing but also reduces the rendering time of a rendering job.