facontidavide / cpp_optimizations_diary

Tips and tricks to optimize your C++ code

About me 👋

Things I love:

• Robotics: when I was a child, I wanted to build robots. This is what I do eventually so... success!
• C++ Programming: I think I am quite good at it. This is called "Dunning Kruger effect".
• Open Source: I have few projects going on, mostly related to robotics. I like reporting bugs and sending PRs titled "Improved performance x%".

cpp_optimizations_diary's Issues

Small vectors - std::array?

Hi Davide, thanks for your excellent writeups!

I was just reading the small vector document, and all the time I was thinking: "He's gonna reveal std::array". But you did not. Have I misunderstood something, or is std::array really a "standard solution" to your StaticVector class?

CPP_Optimizations_Diary/docs/small_vectors.md

Line 107 in e6cca77

## StaticVector and SmallVector in the wild

additional information on c++optimization

As I am also enthusiast about robotics and c++. I found google published abseil along with some of their tips of the week valuable to me.

https://abseil.io/tips/

If you already know it forget what I said.

Additional recommendations

• talking about how amazing vector caching is. I've often been able to get major improvements in speed when just doing something simple like changing an unordered_map to a vector and playing indexing games even with large data. All depends on your L1 cache size.
• Not allocating variables in loops, preallocate any intermediate variables and just assign them in the loop for use. References encouraged when possible.
• While strings are heavy... by god do I never want to go back to the days of using anything else
• Release builds.. I cannot tell you the number of times I've had people complain about performance only to find out they were using debug builds on their production assets.

question about heaptrack roslaunched node

I am trying to use heaptrack to track memory. But how to heaptrack roslaunched executable?

I tried:
heaptrack roslaunch ${pkg_name} ${type_name}

then I found it is heaptracking a python executable "roslaunch" instead of my real executable.

std::deque won't release memory even after shrink_to_fit()

I am using std::deque to receive imu ROS message and odometry ROS message and pop_back() when a lidar message is processed.

It is all working well. But after days of continuous running. The memory consumption goes up to to 90% on my 16 GB memory PC. I used heaptrack to pin down the problem and found std::deque::push_back() allocate 14 GB memory at peak and leak 14 GB memory at the same time. I am not using pointer, the definition of imu_queue is as below:

std::deque<sensor_msgs::Imu> imu_queue_;

I tested my program running for 1 hour. It allocated 1.5GB and leak zero. But the memory consumption do goes up as time went by.

Initially the memory of std::deque is almost none. I checked its size, since I pop_front() all imu message while processing lidar. deque's size is around 37 at most time.

I use shrink_to_fit() after a constant period. But the situation is still the same.

Any clue where did I go wrong?