Fine-grain dataflow model and algorithms for visualization systems

Dataflow computing model is a simple yet powerful mechanism for constructing distributed visualization applications that transform numerical data into images. However, current dataflow visualization systems have serious memory efficiency problems when processing large data sets. In this thesis, we have identified a class of visualization algorithms that require sublinear memory space and are suitable for implementation in these systems. These algorithms are so called fine-grain algorithms, as opposed to the coarse-grain approach adopted by most of the previous systems. The fine-grain algorithms and systems are studied within the framework of fine-grain dataflow computing in general, and the Syntax-directed Dataflow Transformation Method (SDTM) in particular, in this thesis. The SDTM model combines the dataflow programming paradigm with the syntax-directed translation techniques. The new model extends the classical dataflow model in which tokens and nodal functions are atomic. It allows structural definition of the tokens and explicitly states the sequence of nodal actions in relation to the token structures. Unlike previous systems that tend to introduce multiple copies of large data sets during execution, our system is more memory efficient. This efficiency is achieved by keeping a dynamically adjusted minimal window on the input data stream, through the use of attribute grammar to specify attribute dependency and data transformation. Based on the fine-grain algorithms and the SDTM model, we have built a fine-grain visualization system that exhibits faster speed, less memory usage, and higher CPU utilization than a typical coarse-grain system.U of I OnlyETDs are only available to UIUC Users without author permissio