Investigation and prediction of distributed volume rendering performance

Abstract

In this work, I describe the process of developing a cluster scalability model that is capable of predicting performance of a parallel rendering application running on a cluster while only having data that can be obtained from one of its nodes. I begin by studying scaling behavior of a single cluster, employing linear regression and neural networks to construct a cluster-specific scalability model, which im-plicitly captures its hardware characteristics. I use this model as a foundation for further work, developing a hardware-agnostic cluster scalability model. Instead of using explicit hardware characteristics as input, the hardware-agnostic model takes in a distribution of node computation time, which encapsulates local computational load of a rendering application, enabling the model to focus on pre-dicting communication overhead of a cluster. This allows simulation of different hardware by varying the node computation time, gathering enough data to train a neural network that predicts the overall performance of the rendering application on a cluster with arbitrary node hardware.