Probably the most useful observation in the article is hidden right before the Results / Conclusion...
Any optimization should be measured. Otherwise how do you know if all your hard work is worth it? In the author's example, the fastest implementation was actually _not_ lock-free, it merely used a better spinlock than std::mutex.
It's a good idea to use existing implementations mostly so that you can measure and (hopefully without too much trouble) change out the implementation whenever you need to. Different hardware likely has different performance, so measure often!
Any optimization should be measured. Otherwise how do you know if all your hard work is worth it? In the author's example, the fastest implementation was actually _not_ lock-free, it merely used a better spinlock than std::mutex.
It's a good idea to use existing implementations mostly so that you can measure and (hopefully without too much trouble) change out the implementation whenever you need to. Different hardware likely has different performance, so measure often!