A distributed multi-GPU system for high speed electron microscopic tomographic reconstruction
The Howard Hughes Medical Institute and the W.M. Keck Advanced Microscopy Laboratory, Department of Biochemistry and Biophysics, University of California, San Francisco, 600, 16th Street, Room S412D, CA 94158-2517, USA
Ultramicroscopy (March 2011)
@article{zheng2011distributed,
title={A distributed multi-gpu system for high speed electron microscopic tomographic reconstruction},
author={Zheng, S.Q. and Branlund, E. and Kesthelyi, B. and Braunfeld, M.B. and Cheng, Y. and Sedat, J.W. and Agard, D.A.},
journal={Ultramicroscopy},
issn={0304-3991},
year={2011},
publisher={Elsevier}
}
Full resolution electron microscopic tomographic (EMT) reconstruction of large-scale tilt series requires significant computing power. The desire to perform multiple cycles of iterative reconstruction and realignment dramatically increases the pressing need to improve reconstruction performance. This has motivated us to develop a distributed multi-GPU (graphics processing unit) system to provide the required computing power for rapid constrained, iterative reconstructions of very large three-dimensional (3D) volumes. The participating GPUs reconstruct segments of the volume in parallel, and subsequently, the segments are assembled to form the complete 3D volume. Owing to its power and versatility, the CUDA (NVIDIA, USA) platform was selected for GPU implementation of the EMT reconstruction. For a system containing 10 GPUs provided by 5 GTX295 cards, 10 cycles of SIRT reconstruction for a tomogram of 40962×512 voxels from an input tilt series containing 122 projection images of 40962 pixels (single precision float) takes a total of 1845 s of which 1032 s are for computation with the remainder being the system overhead. The same system takes only 39 s total to reconstruct 10242×256 voxels from 122 10242 pixel projections. While the system overhead is non-trivial, performance analysis indicates that adding extra GPUs to the system would lead to steadily enhanced overall performance. Therefore, this system can be easily expanded to generate superior computing power for very large tomographic reconstructions and especially to empower iterative cycles of reconstruction and realignment.
April 15, 2011 by hgpu
