Ricki Blau
EECS Department
University of California, Berkeley
Technical Report No. UCB/CSD-93-736
March 1993
http://www2.eecs.berkeley.edu/Pubs/TechRpts/1993/CSD-93-736.pdf
This dissertation applies performance analysis to the problem of computing complex three-dimensional images. First, it identifies factors that affect the cost of image synthesis and characterizes the complexity of realistic images. Four categories of performance factors are defined: scene characteristics, viewing specifications, rendering parameters, and the computing environment. This classification provides a framework for discussing image complexity and designing performance experiments. The complexity of several complex images from an actual animation workload is described in detail.
A methodology is presented for the construction of reproducible and controllable performance measurement experiments. To measure the performance of a rendering system, an experimenter provides a set of test data, including image specifications. The dissertation describes a portable tool that generates test cases, varying the scene characteristics and viewing specifications under the control of a set of parameters. This model generator, Mg, has been implemented for two different rendering systems. Its test cases have been used to detect performance differences between the two systems and to evaluate the effects of varying the scene characteristics.
Finally, we address the workload partitioning problem for a MIMD rendering system. A simple, low-overhead adaptive algorithm balances the workload effectively on a sixteen-node rendering accelerator. The algorithm uses the rendering time observed for one frame of an animated sequence to predict costs for the next frame. The resulting cost estimates can be used by a second algorithm to divide the work among the available processing nodes. Our cost estimates are approximate, but they are obtained with very little overhead. The net result is an improvement of thirty to eighty percent over the previous load balancing schemes for production-quality rendering of animated sequences. An analysis of several competing schemes demonstrates that tradeoffs between balancing the load and preserving locality are a key consideration in the design of a parallel rendering system.
Advisor: Alan J. Smith
BibTeX citation:
@phdthesis{Blau:CSD-93-736,
Author = {Blau, Ricki},
Title = {Performance Evaluation for Computer Image Synthesis Systems},
School = {EECS Department, University of California, Berkeley},
Year = {1993},
Month = {Mar},
URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/1993/6025.html},
Number = {UCB/CSD-93-736},
Abstract = {This dissertation applies performance analysis to the problem of computing complex three-dimensional images. First, it identifies factors that affect the cost of image synthesis and characterizes the complexity of realistic images. Four categories of performance factors are defined: scene characteristics, viewing specifications, rendering parameters, and the computing environment. This classification provides a framework for discussing image complexity and designing performance experiments. The complexity of several complex images from an actual animation workload is described in detail. <p>A methodology is presented for the construction of reproducible and controllable performance measurement experiments. To measure the performance of a rendering system, an experimenter provides a set of test data, including image specifications. The dissertation describes a portable tool that generates test cases, varying the scene characteristics and viewing specifications under the control of a set of parameters. This model generator, Mg, has been implemented for two different rendering systems. Its test cases have been used to detect performance differences between the two systems and to evaluate the effects of varying the scene characteristics. <p>Finally, we address the workload partitioning problem for a MIMD rendering system. A simple, low-overhead adaptive algorithm balances the workload effectively on a sixteen-node rendering accelerator. The algorithm uses the rendering time observed for one frame of an animated sequence to predict costs for the next frame. The resulting cost estimates can be used by a second algorithm to divide the work among the available processing nodes. Our cost estimates are approximate, but they are obtained with very little overhead. The net result is an improvement of thirty to eighty percent over the previous load balancing schemes for production-quality rendering of animated sequences. An analysis of several competing schemes demonstrates that tradeoffs between balancing the load and preserving locality are a key consideration in the design of a parallel rendering system.}
}
EndNote citation:
%0 Thesis %A Blau, Ricki %T Performance Evaluation for Computer Image Synthesis Systems %I EECS Department, University of California, Berkeley %D 1993 %@ UCB/CSD-93-736 %U http://www2.eecs.berkeley.edu/Pubs/TechRpts/1993/6025.html %F Blau:CSD-93-736