Reduce ROS2 memory usage

asked 2019-02-08 08:37:58 -0600

alsora gravatar image

updated 2019-02-11 04:14:10 -0600

Hi,

I noticed that the memory used by a ROS2 process is really high.

For example running psrecord

$ psrecord --plot listener.png --include-children --duration 10 "ros2 run demo_nodes_cpp listener"

The result will show that a node with only a subscription to a String topic has a memory usage of more than 55 MB. The same amount of memory can be checked using tools like free.

Running the same test with a different node

$ psrecord --plot listener.png --include-children --duration 10 "ros2 run image_tools showimage"

This executable instantiates a subscriber for images, resulting in more than 80 MB of memory.

Are these results consistent with your experience or am I doing something wrong? Is there a way to reduce the ROS2 memory usage?

EDIT: I noticed that part of this memory is due to the execution of the node using the ROS2 CLI, the memory usage reduces when directly calling ./EXECUTABLE_NAME.

EDIT: I have further investigated this issue using valgrind. Thus, without considering the ROS2 CLI process or other things, creating a node with a subscription to an INT64 topic will allocate approx 14 MB. Then I tried creating messages with bigger size (i.e. 0.01 MB, 1 MB, 2 MB, 4MB) and I found a dramatic proportion.

Array4MB.msg

byte[4000000] array

The memory allocated for a node with a subscription is approximately 200 TIMES the dimension of the message. Creating a node which subscribes to a topic where messages of fixed size 4 MB are published, will allocate 800 MB!!

image description

I'm only instantiating the node, it's not spinning or doing anything else.

Thank you

edit retag flag offensive close merge delete

Comments

Try top -H -p PID where PID is the process id of your node- see if there are particular threads that are using the bulk of the memory (though they are all named the same). I see a dozen or more threads per process even when using a single threaded executor, maybe each is taking a bunch of memory

lucasw gravatar image lucasw  ( 2019-02-10 09:58:30 -0600 )edit

I see 5 threads, each using 0.1% MEM. Ubuntu system monitor shows 6.2MB for the process. valgrind and psrecord show 800MB. I trust the second as, if I try to run this node on a device with RAM 512MB it goes out of memory.

alsora gravatar image alsora  ( 2019-02-11 04:02:57 -0600 )edit
lucasw gravatar image lucasw  ( 2019-02-14 10:23:21 -0600 )edit

I was actually going to close this, as what I see from Valgrind for this particular example is completely different from what other tools show. However Valgrind is coherent with other methods when I use small message sizes.

alsora gravatar image alsora  ( 2019-02-14 10:26:01 -0600 )edit

Possibly related - cpu usage is high also https://answers.ros.org/question/3129...

lucasw gravatar image lucasw  ( 2019-02-14 10:26:33 -0600 )edit