Distributing data and control for ray tracing in parallel

Memory sharing for interactive ray tracing on clusters

“…They demonstrated that volume rendering datasets that are too large for the memory of any one computing element is feasible with caching. Badouel et al [6] used a page-based distributed shared memory, similar to one described here, and compared data parallel and image parallel ray tracing programs. They concluded that image parallel rendering with shared memory will scale better than object parallel rendering because of the increased processing and communication overhead that results from more finely dividing the objects in space.…”

Perfect Load Balancing for Demand- Driven Parallel Ray Tracing

“…A load balancing is therefore needed for an efficient parallelization. There are two principal approaches to achieving a balanced load in processors: work stealing [1] and process farming. A problem of both is tuning their parameters.…”

“…Image space subdivision (or screen space subdivision) algorithms [5], [4], [1], [6], [9], [11] exploit the fact that the primary rays sent from the eye through the pixels of the virtual screen are independent of each other. Tracing of primary rays can run in parallel without a communication between processors.…”

“…On the other hand, these algorithms are usually easy to implement. Object space subdivision algorithms [2], [10], [1], [6], [15] geometrically divide the 3D scene into disjunct regions which are distributed in processors' memories. The computation begins with shooting of primary rays into processors storing the regions through which the primary rays pass first.…”

Parallel implementation of ray-tracing algorithm on the Intel Delta parallel computer

“…In many previous studies [6,7,8, 121, each processor was assumed to have access to the whole scene database, so each processor can compute any subimage. Carter et al [8] used a "pixel region module number of processors" scattering decomposition method to achieve static load balancing.…”