Trace: a high-throughput tomographic reconstruction engine for large-scale datasets.

Area of Science:

• Scientific Computing
• Data Analysis
• Imaging Science

Background:

• Modern synchrotron light sources generate massive datasets, necessitating large-scale computation for analysis.
• Computed tomography (CT) is a key imaging technique producing data at tens of gigabytes per second.
• Current CT data reconstruction is computationally intensive, often taking days and hindering scientific progress.

Purpose of the Study:

• To develop a high-performance computing engine for iterative tomographic reconstruction.
• To enable efficient processing of large-scale CT datasets on parallel computers.
• To accelerate scientific discovery by reducing data analysis bottlenecks.

Main Methods:

• Introduced Trace, a data-intensive computing engine for high-performance tomographic reconstruction.
• Implemented fine-grained shared and distributed memory parallelization.
• Utilized a novel replicated reconstruction object data structure with specific optimizations.

Main Results:

• Achieved a 158x speedup using 32 compute nodes (384 cores) compared to a single-core setup.
• Reduced the reconstruction time for a large sinogram from 12.5 hours to under 5 minutes per iteration.
• Demonstrated significant performance gains on datasets from the Advanced Photon Source.

Conclusions:

• The Trace engine efficiently processes large-scale tomographic data.
• The parallelization and optimization techniques minimize reconstruction times.
• Accelerated CT data analysis facilitates scientific progress.