NVlabs
Significant Increase in Scene Loading Time from Sionna 0.16.2 to 0.18
I have noticed a significant increase in the amount of time it takes to load a scene from a '.xml' file exported from Blender using Mitsuba after upgrading from Sionna version v0.16.2 to v0.18. The same '.xml' file that took 5 seconds to load now takes around 8 minutes in the new version. Also, the RAM usage exceeds 40 GB, which seems unusually high.
Tested versions: Sionna 0.16.2 | tensorflow 2.15.0 | python 3.9.19 Sionna 0.18 | tensorflow 2.15.0 | python 3.9.19
Ubuntu 22.04.4
Is this increase in loading time and memory usage a known issue? Could it be related to the new intelligent surfaces introduced in the latest version?
Hi @kjellwagn, Does this behavior only occur when loading your custom scene, or do you also experience it with the built-in scenes?
Is your custom scene particularly large or does it contain a very detailed mesh?
Hey @SebastianCa ,
This also occurs with the built-in scenes, although the time increase is minimal. For example, the Munich scene takes an average of 5 seconds to load in v0.16.2 and 7.5 seconds in v0.18. In fact, the custom scene being used is quite large and detailed, comprising about 5.5 million triangles.
That sounds like a fairly large scene. We’ll look into it and get back to you.
Hello @kjellwagn,
We have identified some performance regressions, and fixes will be included in the next release. Thanks for bringing it to our attention!
In terms of memory usage, I didn't measure a particular difference on the built-in Munich scene. Could you please share the scene where you've measured increased usage? Please include all meshes and the XML file.
I appreciate your support. Due to company restrictions, I can't share the entire scene, so I have attached a cutout. However, one can still notice the performance penalty regarding loading time and memory usage. cutout.zip
Thanks for sharing the scene.
I have done the following measurements with nvtop for memory usage & runtime:
They all reach the same memory usage, mostly due to TensorFlow allocating a large memory arena from the start. How did you measurement memory usage?
On the CPU I have checked and the upcoming version seems to use roughly as much memory as v0.16.2, so hopefully the loading time fix also significantly decreased memory usage.
