CPU multi-core questions

[paulm]

Hi,

I'm trying to debug some performance issues between different nodes. I'm seeing the camera alignment phase of the Align Photos step only using 1 CPU core (i.e. no mult-threaded and/or GPU computations). It this expected behaviour? I can't seem to find any documentation on this topic. This is with 1.4.2 Pro under Linux.

Secondly, are there any performance limitations in demo mode? Due to having only a single professional license I am running one test in activated mode on one node and another in demo mode on a different node.

Regards,
Paul

[Alexey Pasumansky]

Hello Paul,

Which sub-step of Align Photos stage you are speaking about: feature points detection, preselection, image matching, scene structure estimation? Also what is the size of the dataset and processing parameters used?

Actually, all these stages are well parallelized, so I can only assume that the dataset is too small, so you are not able to see the multithreading.

[paulm]

The step is "Estimating camera locations", sorry for misremebering the name. The scene has 199 photos of 4000x3000 pixels, also containing 25 markers, no masks. Settings: accuracy = high, generic preselection enabled, reference preselection enabled, 40k key point limit, 4k tie point limit, adaptive camera model fitting.

The highest I'm seeing is around 108% CPU out of a possible 1600% (i.e. 16 cores). The node has 32 GB RAM and also 2 Tesla K40m's.

[paulm]

So I did some further testing and it seems I DO get multi-threaded computations when processing the same dataset on my laptop (4 cores with HT + Quadro M1000M), while on a more powerful compute node (16 cores + 2x Tesla K40m) I once again see the single-core computations. I use a Python script to make sure I'm doing the exact same processing on these two nodes and the single-core part is due to chunk.alignCameras().

The matchPhotos() step is roughly 2.5 times faster on the compute node, but after that performance switches to single-core.

[Alexey Pasumansky]

Hello paulm,

Can you share the project in PSZ format saved after the processing (with just alignment results and markers) with support@agisoft.com, so that we can try to reproduce the same behavior basing on the matching points that are already in the project?

[paulm]

Hi Alexey,

I just sent the dataset as a .psz file. This contains the results of doing Workflow -> Align Photos from the GUI. The "Estimating camera locations" step used only 1 CPU core in this case.

Paul