Performance consideration

Reconstruction

Currently, the bottleneck of the data processing pipeline is the preprocessing part, which is still implemented in CPU. The total number of threads of the current benchmark machine is only 32. We expect a significant performance boost on a GPU node with a decent number of threads.

The 3 GB/s throughput was measured using a projection size of 512 x 512 pixels and 500 projections per scan. The throughput will decrease as the total size of data per scan increases since the performance (in particular the preprocessing part on CPU) does not scale linearly. If you need a high throughput with large projections sizes and/or large numbers of projections per scan, it is recommended to down-sample the projections, which can be configured in the GUI.

Visualization

The data rate in RECASTX is really high not only in terms of the raw projection images but also in terms of the reconstructed images and volumes. In the default setup, the size of the reconstructed volume is 128 x 128 x 128 pixels, which amounts to 8 MB. Suppose that you would like it to update at 5 Hz, the data rate will be 40 MB/s. It should be noted that the high-resolution slices have not been included. Even though, you may laugh off it because you already have a 10 Gbps network connection between your client and server machines in your beamline or lab. However, this could be a problem if you are going to conduct a remote experiment and connect your GUI client to the reconstruct server via SSH.

Even with a 10 Gbps network connection between the client and the sever, the network bandwidth could be a bottleneck if you want to increase the resolution of the reconstructed volume. A high-resolution volume of 1024 x 1024 x 1024 pixels amounts to 4 GB and would be difficult to update at just 1Hz (unless you compress data) if you don't have a 100 Gbps network connection between the client and the server. Therefore, you should not rule out the possibility of running the GUI client and the reconstruction server on the same machine if it is powerful enough.