Nabu latest updates from the last two months.

1. Current status and capabilities
2. Latest updates
3. Future plans

Nabu is a versatile tomography reconstruction software.

• It offers a set of high performance building blocks
• On top of them, reconstruction pipelines are built for specific use cases:
  • Full-field absorption/phase contrast (now)
  • Diffraction tomography (work in progress)
  • Fluorescence tomography (in the future)
• Extensive documentation

Support for absorption and phase contrast tomography comes with two pipelines.

Chunked Pipeline processes slabs of data volume.

It is fast and adapted to usual settings.

Grouped Pipeline processes groups of radios.

It is slower but addresses the shortcomings of the previous one.
Use case:

• Big datasets (eg. large number of projections)
• and
• Phase contrast with a large filter (CTF, Paganin with a large delta/beta)

Versatile

• Absorption & phase contrast (now), diffraction (in progress), fluorescence (in the future)
• Save checkpoints and resume the processing from a saved checkpoint
• Can plug with other libraries: pyFAI (XRD-CT), pyMCA (XRF-CT)
• Comes with a variety of tools:
  • Alignment (used by Bliss EBS-tomo)
  • Detector tilt and center of rotation estimation

Fast

• All steps (except one option, see next slides) are available on GPU
• Full volume from 1200 projections (no HA): 2 mins
• Full volume from 7000 projections (no HA): 8 mins
• Full volume from 4400 projections (HA): 9 mins
• Reconstruction speed was limited by network data access - much better since GPFS deployment (still ~50% of the time in I/O)

• Add GPU support for Contrast-Transfer-Function phase contrast pipeline
  • Tests with A. Pacureanu & P. Cloetens
  • ⚠️ Distortion correction not not available on GPU

More tests to be done with ID16A.

• Some XRD acquisitions can take a couple of days ; by then it's too late to start integrating the data.
• Need to perform azimuthal integration as soon (and fast) as possible
  • Since GPFS deployment, acquisition/file transfer is no longer the bottleneck
  • pyFAI cannot cope with the acquisition pace if using a single device

→ Distribute the azimuthal integration over the computing cluster

This addresses ID15A needs, but there is also interest from ID11.

Work in progress: distributed azimuthal integrator

• gitlab.esrf.fr/tomotools/integrator
• Automatically distribute the workload on the SLURM cluster
• "Online scan watcher/integrator": integrate data as soon as available
• Determining the most suited machines for distributed integration (GPU or CPU)

This solution scales nicely and leverages both GPFS and computing power.

Overall Frames per Seconds (FPS) on a ID15 dataset containing 25 000 images

• Use distributed AI on large collection of datasets
• Tests the "online scan integrator" on ID15A
• Start XRF-CT support