Audit code for "false sharing" anti-patterns

[mcalancea]

False sharing came up during conversation on #624. There is no confirmed instance of it happening yet (just one strong hunch 😄), but examining the code-base from this perspective may be advisable.

[hero78119]

I just add 2 incremental commits on top of master branch, aiming to narrow down the issue

avoid in-place change

with commit 50f1213

• same way of parallel traverse
• avoid in-place vector change

And here before/after this change on fibonacci 2^20

fibonacci_max_steps_1048576/prove_fibonacci/fibonacci_max_steps_1048576
                        time:   [4.7952 s 4.8269 s 4.8525 s]
                        change: [-2.7392% -0.9628% +0.7617%] (p = 0.33 > 0.05)
                        No change in performance detected.

benchmark result shows identical, which imply false sharing is not the root cause for performance degrade

change parallel traverse

on top of prev commit, add new commit 269c9fa

• move traverse parallism to higher level

And here before/after this change on fibonacci 2^20

fibonacci_max_steps_1048576/prove_fibonacci/fibonacci_max_steps_1048576
                        time:   [4.1463 s 4.1676 s 4.1861 s]
                        change: [-15.930% -14.492% -13.011%] (p = 0.00 < 0.05)
                        Performance has improved.

Both experiment indicate false sharing probably not being the root cause, whereas the way of traverse is.