Memory usage keeps increasing during large-scale mapping – eventually crashes at \~98%

[mahmood-ul-hasan]

Thank you for your work on this project — it's been very helpful.
I'm using ros2 glim_rosnode to build a large-scale map. After a certain point, memory usage steadily increases until CPU memory reaches approximately 98%, after which the node crashes.
The following memory warnings repeatedly occur:

[glim_rosnode-1] [2025-05-30 14:57:35.294] [mem] [warning] CPU memory usage: 28979.05 / 31497.23 MB 92.01%
[glim_rosnode-1] [2025-05-30 14:57:40.294] [mem] [warning] CPU memory usage: 28938.97 / 31497.23 MB 91.88%
[glim_rosnode-1] [2025-05-30 14:57:50.295] [mem] [warning] CPU memory usage: 29336.95 / 31497.23 MB 93.14%

Eventually, it reaches >98% and the node crashes due to memory exhaustion.

Questions:

1. Are there any parameters (e.g., cache size, voxel resolution, GPU memory use options) that can be adjusted to reduce memory usage or offload more to the GPU?
2. Is there a recommended strategy for long-duration mapping with high memory efficiency?
3. Is memory usage expected to grow indefinitely, or should it plateau during normal operation?

Any suggestions would be very helpful.
Thanks in advance!

[koide3]

1. Please try tuning the following parameters:
   • Decrease "submap_target_num_points" in "config_sub_mapping_gpu.json" to 20000 ~ 30000 (decrease # points per submap)
   • Decrease "keyframe_voxelmap_levels" in "config_sub_mapping_gpu.json" to 1 (decrease # layers of voxelmaps per submap)
   • Increase "max_implicit_loop_overlap" in "config_global_mapping_gpu.json" to 0.3 ~ 0.4 (reduce # of factors in global optimization)
2. If only the CPU-RAM matters, I recommend setting 16 ~ 32 GB swap memory to mitigate the problem https://www.digitalocean.com/community/tutorials/how-to-add-swap-space-on-ubuntu-20-04
3. The memory usage grows indefinitely as the map area grows

[mahmood-ul-hasan]

@koide3 Thank you so much for your support and help.

[koide3]

I implemented GPU memory offloading that enables swapping out submap data from the GPU memory and drastically reducing the memory consumption. You may want to try it: github.com//pull/235

[TokyoWarfare]

I recently updated to the latest and I could not understand why memory still kept growing

I realized the the PR is not merged :D

I'll test the "offload" branch ASAP and report back. Cheers!