COZ: Finding Code that Counts with Causal Profiling
This paper is worth reading just for one neat idea. Everyone wants to make their code fast. If you had a magic wand that enabled you to ‘make function X’ faster then you could measure the impact on the performance of the system and you would know how important said function is to the overall performance. Obviously, this is a bit silly because if you had such a wand, you would just use it everywhere and not have to optimize the code yourself.
This paper inverts this idea. Instead of making function X faster, they leverage a neat trick to make every other function slower. Then the relative importance of the function to overall system performance can be measured. The results are impressive. Neat!
Note: The paper doesn’t just look at whole functions.
