Video on Measuring / Increasing Performance of Code

Emery Berger talks about performance improvements that work / don't work.

Key Points

When optimizing your code, you can easily get random side effects which impact performance more than the changes.
The execution environment (command args, current directory, environment variables) can have significant impact on execuiton speed, so they have to be factored out when measuring performance.
Use stabilizer to randomize the layout of a linked binary and ignore accidental sideeffects.
-O3 has no statistically significant benefit over -O2 in clang / LLVM.
Typical profilers come from the single threaded single core time, and are not well suited to profile async or concurrent code.
Use coz-profiler to find code that causes performance problems.
Users don't care about the performance they care about responsiveness. This may or may not be the same thing.