Метод збалансованого завантаження рендерів при формуванні тривимірних графічних сцен

Abstract

This paper addresses the problem of load balancing in the process of rendering three-dimensional graphical scenes, which remains a critical challenge in high-performance computer graphics systems. It is shown that traditional static and reactive load distribution approaches are not sufficiently effective under conditions of heterogeneous computational complexity within a frame. In modern rendering pipelines, different regions of an image may significantly differ in geometric density, depth complexity, shader cost, transparency, and presence of computationally expensive visual effects such as reflections, shadows, and ray tracing. As a result, uniform spatial partitioning of the frame does not guarantee balanced workload distribution among rendering nodes. To overcome these limitations, a method of predictive-adaptive tile distribution is proposed. The method is based on multicriteria evaluation of tile complexity, which integrates geometric characteristics, fragment processing cost, shader complexity, transparency factors, expensive visual effects, and historical performance data from previous frames. Each tile is assigned an integral weight reflecting its predicted computational cost. Tiles are then sorted in descending order of complexity and distributed among rendering nodes according to the criterion of minimizing the current normalized workload. The predictive-adaptive nature of the method is ensured by incorporating a historical component that updates tile complexity estimates based on previous rendering iterations. This enables the system to exploit temporal coherence in dynamic scenes and improve the accuracy of workload prediction over time. Unlike purely reactive approaches, the proposed method prevents load imbalance before execution, reducing synchronization overhead and improving frame time stability. The method is applicable to tile-based, cluster, and multi-GPU rendering systems and can be effectively used in both homogeneous and heterogeneous computing environments. Experimental estimation shows that the computational overhead of the proposed approach is acceptable relative to the overall frame rendering time, while providing improved load balancing and resource utilization. The proposed approach demonstrates higher efficiency compared to static, reactive, and single-criterion cost-based methods, particularly in complex and highly non-uniform rendering scenarios.

У статті розглянуто питання балансування навантаження в процесі рендерингу тривимірних графічних сцен. Показано, що традиційні статичні та реактивні підходи не забезпечують достатньої ефективності в умовах неоднорідного розподілу обчислювальної складності кадру. Запропоновано метод прогностично-адаптивного розподілу тайлів між рендер-вузлами, який ґрунтується на багатокритеріальному оцінюванні складності тайлів. Метод може бути використаний у системах тайлового, кластерного та multi-GPU рендерингу.