Analytical Performance Prediction for Iterative Reconstruction Techniques in Electron Tomography of Biological Structures

Abstract

The problem of the three-dimensional reconstruction of structures from projection data occurs in a wide range of areas. Algebraic reconstruction techniques (ART) are iterative procedures for recovering the structure of three-dimensional objects from projection images. ART was dismissed during the seventies due to the high demands on computing resources. In the following years, computer algorithms for image reconstruction with fast convergence were designed, implemented, evaluated, and, finally, optimized. Nevertheless, the reported experimental data exceed computing potential for a single CPU as well as I/O capability. Interesting recent research aims at acquiring experience with parallelization strategies and at demonstrating the effectiveness of the massively parallel processing approach in biological environments. This makes the investigation of performance prediction for high-performance parallel computing of paramount importance. A performance prediction model for iterative reconstruction techniques (IRT) would provide additional knowledge of the parallel computer algorithm and predict its behavior under specific conditions or hardware configurations. Also, a goal of performance analysis is to find the set of parameter values that produces the best overall performance. Not only the end-users of these systems have a vested interest in predicting performance but also the computer designers and the professional engineers. This article addresses the derivation and evaluation of an analytical performance prediction model for a parallelization of IRT (block iterative version of the component averaging, BICAV) and emphasizes the essential role of parallel computers in image reconstruction. The techniques’ behavior is analyzed through specifics steps to create an analytical formulation of the problem. BPTomo is a parallel distributed application for tomographic reconstruction that uses IRT. The analytical performance prediction model is validated by comparison of the estimated times for representative datasets against BPTomo computation times measured on two PC clusters. The analytical model is shown to be quite accurate. The percentage deviation between estimated and measured times is less than 12%.