D435 High CPU Usage

[cynosure4sure]

Required Info
Camera Model	D435
Firmware Version	05.08.15.00
Operating System & Version	Linux (Ubuntu 14.04
Kernel Version (Linux Only)	4.4.0.111
Platform	PC
SDK Version	LibRealSense v2.9.1

Issue Description

I am using D435 camera in Ubuntu 14.04 with ROS.
For only Color Camera and Depth Camera enabled I get a high CPU usage on my i7-6700HQ laptop consistently 50% across all threads. Please see the image below for only Color and Depth Camera running with pointcloud disabled.
It seems very unusual as compared to using ZR300 module.

[fgzhu]

turn of OpenMP default will solve this problem

[cynosure4sure]

Can you share your CPU usage after turning off OpenMP?

[fgzhu]

less than 5% if just grab frame without showing

[cynosure4sure]

@fgzhu Thank you for sharing your CPU usage.
After compiling with OpenMP=OFF and only grabbing color and depth frames without visualization, I am still seeing CPU usage of 50-60% on one thread as shown below. Are there any other options that you turned off and can you please let me know the specs of your system. Thank you.

[zivsha]

@cynosure4sure do you use a pre-compiled version of librealsense? or are you building both the wrapper and the library?

[cynosure4sure]

@zivsha I am building both librealsense and ros wrapper from source.
Now my CPU usage is below 10% on single thread.
I compiled librealsense with the following flags:
cmake ../ -DBUILD_EXAMPLES=OFF -DBUILD_GRAPHICAL_EXAMPLES=OFF -DBUILD_WITH_OPENMP=OFF -DBUILD_UNIT_TESTS=OFF -DCMAKE_BUILD_TYPE=Release -DENABLE_ZERO_COPY=ON -DFORCE_LIBUVC=OFF

and the ROS wrapper -DBUILD_WITH_OPENMP=OFF

I also have turned off the depth/color alignment. So now I am happy with the reasonable CPU usage. If you have any other advice for further improvement that would be helpful.

One more thing, if I turn on depth/color image alignment my CPU usage still jumps to 50% single thread.

[mfe7]

I also followed this suggestion to disable OpenMP, and I see the CPU usage is very low when just grabbing frames without visualizing (<5% of a core). When I also enable pointcloud computation, the CPU usage hardly changes. As soon as I start visualizing the pointcloud, the CPU usage rises to >100%. That is not surprising, because I assume visualization is intensive.

The strange thing is that when I do rostopic hz <pointcloud_topic>, the CPU usage reaches >100% as well. So, it seems like the realsense node doesn't do much until something subscribes to one of its topics, and then uses a lot of CPU.

Is that expected behavior? Or is there something I am doing incorrectly?

[RealSense-Customer-Engineering]

[Realsense Customer Engineering Team Comment]
hi @mfe7,

we have some new implementation for better CPU utilization, can you please help to confirm if latest v2.14.0 release still have this issue?
thanks.

[RealSense-Customer-Engineering]

[Realsense Customer Engineering Team Comment]
hi @mfe7,

is it works for you?

[mfe7]

@RealSense-Customer-Engineering no, my issue is not resolved.

I just installed the latest firmware on the D435 (5.11.1), got the latest Debian librealsense pkgs, and compiled the latest realsense_ros version.

Here's my htop when running roslaunch realsense2_camera rs_camera.launch filters:=pointcloud, and nothing is subscribing to the topics (10% of a CPU):

When I do rostopic echo on the /camera/color/image_raw and /camera/depth/color/points topics (41% of a CPU):

When I view those two topics in RViz (113% of a CPU):

[mfe7]

I just tested the same camera on a different computer (the above comment refers to a laptop Gigabyte Aero 15", this comment uses a desktop), with the same installation setup (debian librealsense, built latest realsense ros pkg).

The CPU usage is much more reasonable on the desktop, with the ROS nodelet manager capping out at <50%CPU when viewing the pointcloud in RViz. When I use a 424x240 depth and color stream, it is consistently <30%CPU.

It might be an issue specific to the laptop I have been using, but it is a reasonably powerful gaming laptop so that is surprising.