Humble -> Jazzy Command-/StateInterface Migration: 10x slowdown

[mhubii]

Hi All,

Just curious to learn about the new framework-managed interfaces. It seems Franka / Universal have not adopted this change yet.

Are there any benchmarks for the new std::unordered_map? I just tried updating the lbr_fri_ros2_stack, and approximately experience a 10x slowdown.

Also refer migration guide: https://control.ros.org/jazzy/doc/ros2_control/doc/migration.html#migration-of-command-stateinterfaces.

Thank you for any insights.

[mhubii]

Just for an example, reads that used to be (inside hardware_interface::SystemInterface::read)

hw_state_ = some_robot_driver->get_state();

now expand to multiple lookups

auto state = some_robot_driver->get_state();

set_state("SOME_ARRAY_0", state.some_array[0]);
set_state("SOME_ARRAY_1", state.some_array[1]);
set_state("SOME_ARRAY_2", state.some_array[2]);
set_state("SOME_SCALAR", state.some_scalar);
// ... 

hence the slowdown. Is this a user mistake?

[christophfroehlich]

No, your code snippet seems to be fine.

I'd expect some overhead with the unordered_map, but 10x? I'm not sure.. How have you made this measurement? Using the published statistics, read/write execution times? Can you provide a screenshot from before/after the change?

[mhubii]

My benchmark prints to the terminal 😅

auto start = std::chrono::high_resolution_clock::now();
// do stuff...
auto end = std::chrono::high_resolution_clock::now();
std::chrono::duration<double, std::milli> read_duration = end - start;
RCLCPP_INFO_STREAM(get_node()->get_logger(),
                   "SystemInterface::read() took " << read_duration.count() << " ms");

Old lookup (running on Jazzy)

- name: lbr_system_interface.stats/read_cycle/execution_time/average
  value: 1.7262097799971798
[...]
- name: lbr_system_interface.stats/write_cycle/execution_time/average
  value: 0.41329963252614654

New lookup (running on Jazzy)

- name: lbr_system_interface.stats/read_cycle/execution_time/average
  value: 19.53063345686052
[...]
- name: lbr_system_interface.stats/write_cycle/execution_time/average
  value: 4.728137644319869

[christophfroehlich]

you can plot the execution time from the statistics topic, see
https://control.ros.org/jazzy/doc/ros2_control/doc/introspection.html

for reference, how many state interfaces do you have in your component?

[mhubii]

for reference, how many state interfaces do you have in your component?

67 state; 20 command

[saikishor]

for reference, how many state interfaces do you have in your component?

67 state; 20 command

@mhubii did you compile with release flags?

[mhubii]

for reference, how many state interfaces do you have in your component?

67 state; 20 command

@mhubii did you compile with release flags?

yes of course: https://github.com/lbr-stack/lbr_fri_ros2_stack/blob/56d0828d9a3695bfbd1cbe6790c446dfc62b1d08/lbr_fri_ros2/CMakeLists.txt#L8

[mhubii]

don't get me wrong, this still works fine, it just appears to be significantly slower than before

[saikishor]

don't get me wrong, this still works fine, it just appears to be significantly slower than before

No no totally fine, then let me check and see what we can do. I faced this issue and it got fixed with Release mode, that's why I asked you. However, I did with --mixin flag of colcon build. It should be exactly the same.

[mhubii]

Would you recommend adopting this change early regardless? Since export_command/state_interfaces is gonna get deprecated

ros2_control/hardware_interface/include/hardware_interface/hardware_component_interface.hpp

Line 359 in eba8665

[[deprecated(

I think conceptually it is much simpler now, just slower unless I find a fix here.