Speaker
Description
HTTomo stands for High Throughput Tomography pipeline for processing and reconstruction of parallel-beam tomography data. The HTTomo project was initiated in 2022 at Diamond Light source in anticipation of major data increase with the Diamond-II upgrade. With the support of modern developments in the field of High Performance Computing and multi-GPU processing, the main goal is to mitigate I/O bottlenecks and through GPU acceleration enable higher throughput for big data.
HTTomo is a user interface (UI) written in Python for parallel data processing using MPI protocols. It orchestrates I/O data operations and enables processing on a CPU and/or a GPU using computing cluster or a personal computer. HTTomo utilises other data analysis libraries, such as TomoPy for CPU and HTTomolibgpu for GPU, as backends for data processing. The methods from the libraries are exposed through YAML templates to enable fast task programming and pipeline building.
The main concept of HTTomo is to split the data optimally in accordance to the given computational resources and methods requirements. It is a GPU-memory aware system that exploits the GPU device in a way, that the data stays, for as long as possible, on the device. We use CuPy's API to help with the data transfers and faster computations.
In the presentation, the main elements of HTTomo's framework will be demonstrated, as well as, the benchmarks with the current tomographic software in production at DLS, Savu.
| Abstract publication | I agree that the abstract will be published on the web site |
|---|
