On the design and performance of the PARFUM Parallel Fault Tolerant Volume Renderer

Abstract

Volume rendering is an important and CPU-intensive technique for visualizing large scalar fields. In essence, a volume renderer performs two activites on behalf of the user: loading a new data set, and rendering the current one. At one level, the performance of an individual activity is important. At another level, the erformance of the session as a whole, in particular switching from one activity to the next, should be addressed. In this paper we present PARFUM, a parallel volume renderer based on a controller/worker model in a network of workstations. PARFUM has three essential properties that increase the performance of a parallel volume renderer. First, dynamic load balancing is employed during a rendering activity. Second, workers may enter or fail without affecting the correctness of a session. Third, a user may easily abort the current activity in favour of a new one. These properties may more easily be achieved by accepting (rather than fighting) the inherent asynchrony in a distributed system. As a consequence, PARFUM attempts to minimize causal dependencies in the interaction between the user and the controller as well as between the controller and the workers. We evaluate two implementations based on the TCP and UDP transport protocols respectively.