high performance computing on graphics processing units: hgpu.org

In interactive volume rendering, the cost for rendering a certain block of the volume strongly varies with dynamically changing parameters (most notably the camera position and orientation). In distributed environments – wherein each compute device renders one block – this potentially causes severe load-imbalance. Balancing the load usually induces costly data transfers causing critical rendering delays. In cases in which the sum of memory of all devices substantially exceeds the size of the data set, transfers can be reduced by storing data redundantly. We propose to partition the volume into many equally sized bricks and redundantly save them on different compute devices with the goal of being able to achieve evenly balanced load without any data transfers. The bricks assigned to a device are widely scattered throughout the volume. This minimizes the dependency on the view parameters, as the distribution of relatively cheap and expensive bricks stays roughly the same for most camera configurations. This again enables our fast and simple scheduler to evenly balance the load in almost any situation. In scenarios in which only very few bricks constitute the majority of the overall cost a brick can also be partitioned further and rendered by multiple devices.