Speaker
Description
The gradual shift to 4th-generation synchrotron sources has been boosting data production rates that nowadays easily exceed 1-10 TB of produced raw image data per day. This has not only created a big data bottleneck in terms of data storage, management, processing, and visualization, but also requires a revised approach in utilizing state-of-the-art data processing techniques. In the recent past, several open-source efforts have been established, ranging from reconstruction algorithms [1-3], volume data analysis and visualization tools [4] to storage and metadata handling frameworks [5]. However, these can be considered “silo” solutions that not necessarily interface with each other smoothly. Moreover, in many cases the transition from single proof-of-concept-based code toward making software available to a wider (big) imaging community is still far from being established.
Both at TOMCAT and at LNLS, we have independently been developing and utilizing state-of-the art processing tools. As one of the biggest data producers of the Swiss Light Source, the TOMCAT beamline has guaranteed smooth user operation for over a decade now thanks to an efficient data pipeline [6]. More recently, real-time reconstruction capabilities [7] and TB-sized volume analysis tools [8] have been added to ease the data analysis challenge. At LNLS, we have pioneered novel GPU-enhanced tomographic reconstruction algorithms [9], including phase recovery filters in both directions [10]. Furthermore, we have advanced in-memory processing capabilities, allowing visualization with optimized and fast rendering using the Nvidia/Index API [11] as well as on-the-fly segmentation [12], both using the RDMA protocol. In summary, all these strategies represent cornerstones that can now be integrated into a generalized platform as well as conventional graphical interfaces that will further allow the development of user-defined plugins and facilitate rapid exchange.
In the present work, we devise and describe a general architectural concept for data flow in typical tomographic imaging experiments, which involves both the underlying application stack with its “gluing” components as well as a full operating model in a standardized HPC environment. We present and discuss the feasibility of tomography scans with processing times of several seconds up to a minute, for volumes of several hundreds of GBs. To achieve stable operation, we leverage recent developments in IT architectural frameworks and apply industry-standard best practices for modularity, virtualization, and CI/CD. We show how our architecture drastically improves the user experience. Finally, we discuss different implementations of GPU/CPU communication and present benchmarks of reconstruction and visualization tools using different hardware.
References:
[1] D. Gürsoy, F. De Carlo, X. Xiao et al., J. of Synchrotron Radiat 21, 1188, 2014.
[2] V. Nikitin, J. Synchrotron Radiat 30, 179, 2023.
[3] W. van Aarle et al., Opt. Express 24(22), 25129, 2016.
[4] A. Aboulhassan et al., J. Imaging 8(7), 187, 2022.
[5] Moore, J., Allan, C., Besson et al., Nat. Methods 18(12), 1496, 2021.
[6] F. Marone, A. Studer, H. Billich et al., Adv. Struct. Chem. Imaging. 3(1), 1, 2017.
[7] J.-W. Buurlage et al., Sci. Rep. 9(1), 18379, 2019.
[8] A. Miettinen, I. V. Oikonomidis, A. Bonnin et al., Bioinformatics 35(24), 5290, 2019.
[9] E. X. Miqueles et al., PPSC 2020. https://doi.org/10.1137/1.9781611976137.3
[10] E. X. Miqueles, P. Guerrero, Results Appl. Math. 6, 100088, 2020.
[11] T. V. Spina et al., JACoW 2021 , https://doi.org/10.18429/JACoW-ICALEPCS2021-FRBL05
[12] A. Pinto et al., Synchrotron Radiat. News 35(4), 36, 2022.
| I plan to submit also conference proceedings | No |
|---|
