Superposition on a multicomputer system

• Published in: Medical physics (Lancaster) Vol. 18; no. 3; p. 468
• Main Authors: Murray, D C, Hoban, P W, Round, W H, Graham, I D, Metcalfe, P E
• Format: Journal Article
• Language: English
• Published: United States 01.05.1991
• Subjects: Algorithms; Computer Systems; Humans; Radiotherapy Planning, Computer-Assisted - methods
• ISSN: 0094-2405; 2473-4209
• DOI: 10.1118/1.596650

Superposition (convolution using a noninvariant kernel) has been shown to be a highly promising technique for use in calculating dose distributions in radiotherapy treatment planning. However, one major difficulty that currently prevents use in routine planning is the computational effort required to perform the calculation in three dimensions. To help solve this problem the superposition technique has been implemented on a parallel processor multicomputer in order to examine the performance characteristics of such a system. Up to eight elements have been connected in a pipeline (linear array), and tree networks of three and seven processors have also been constructed (using INMOS T800 transputers). The significant results obtained with these networks are: (1) Both topologies provide near-linear speedup with increasing processor number (8 processors provide 7.81 times the computing power of a single processor when using an optimal communication packet size); (2) increasing communication packet size from 1 voxel to an optimum of approximately 40 voxels significantly reduces communication overhead per processor. Overhead per processor for a 7-element linear array is 6.9% when using 1-voxel packets, but only 1.8% when using 40-voxel packets; (3) the topology of the network has some effect on communication overhead: Arranging 7 processors in a 1-2-4 binary tree reduces overhead to 80.1% of that encountered using a 7-element linear array (with packet size of 1 voxel).