fujitatomoya
commented
10 months ago export ROS_DISABLE_LOANED_MESSAGES=0 this needs to be done, since https://github.com/ros2/rcl/pull/1110
some experimental check... using docker container.
- binding
/dev/shmis the only requirement to use shared memory transport and data sharing provided by Fast-DDS
tomoyafujita@~/DVT >docker run -it --network host -v /dev/shm:/dev/shm tomoyafujita/ros:rolling
root@tomoyafujita:~# ls -lt /dev/shm/
total 0
root@tomoyafujita:~# ros2 daemon status
The daemon is not running
root@tomoyafujita:/# ros2 run demo_nodes_cpp talker_loaned_message
[INFO] [1703119123.500535244] [loaned_message_talker]: Publishing: '1.000000'
[INFO] [1703119123.500835352] [rclcpp]: Currently used middleware can't loan messages. Local allocator will be used.
[INFO] [1703119123.500871185] [loaned_message_talker]: Publishing: 'Hello World: 1'
[INFO] [1703119124.500472323] [loaned_message_talker]: Publishing: '2.000000'
[INFO] [1703119124.500608551] [loaned_message_talker]: Publishing: 'Hello World: 2'
...- what if only host network interface is shared? (host system case does not have this case, since everyone should have
/dev/shmin the same namespace.)
tomoyafujita@~/DVT >docker run -it --network host -v /dev/shm:/dev/shm tomoyafujita/ros:rolling
root@tomoyafujita:/# ros2 node list
root@tomoyafujita:/# ros2 topic list
/chatter
/chatter_pod
/parameter_events
/rosoutinteresting ths case, we cannot see the nodes but topic list. i think Simple Participant Discovery Protocol (SPDP) uses shm transport, so that participants cannot discover each other. on the other hand, Simple Endpoint Discovery Protocol (SEDP) falls back to multicast based default discovery, that can discover the DataReader and DataWriter, so we can see the topics.
https://github.com/eProsima/Fast-DDS/pull/3753 related to this behavior, see more details for https://fast-dds.docs.eprosima.com/en/latest/fastdds/discovery/simple.html#simple-discovery-settings
Note:
root@da343f3b6b7d:/# fastdds shm clean
shm.clean:
0 ports in use
0 segments in use
3 zombie ports cleaned
2 zombie segments cleaned
I had this question during ROSCon 2023 presentation. I believe that this is one of the use case to show how we can use the data sharing zero copy even with container and Kubernetes. Certain constraints that we need to assign the containers in the same pod as localhost, but for the explanation this makes sense to add the example here.