Audit atomics and use weakest ordering possible #6477
Assignees
Labels
domain: performance
Anything related to Vector's performance
type: task
Generic non-code related tasks
Comments
blt
added
type: task
This was referenced Feb 17, 2021
blt
added a commit
that referenced
this issue
Feb 19, 2021
…6480) * Adjust the strength of `gauge_add` atomics This commit adjusts the strength of the `gauge_add` atomics used by `OpenGauge`. Previous implementation called `emitter` potentially `n^2 / 2` times, where `n` is the number of participants. Current implementation calls `emitter` only once after the participant CPU manages to adjust `OpenGauge` value, though no guarantee is made about the ordering of participant calls to `emitter`. That is, this commit guarantees that thread `A` will sequence its call to `emitter` after modification of the internal `gauge` but makes no guarantee about sequence of thread `B`'s parallel call to its `emitter`, relative to `A`. If the point of this code is to sequence calls to `emitter` between threads we can close this out and I'll make a different PR to clarify in the documentation. Closes #6477 Signed-off-by: Brian L. Troutwine <[email protected]> * Update documentation, introduce spin_loop hint After clarification in review of the previous commit it was clear that this code is intended to implement a spinloop with `emitter` in the critical section. Update this code's documentation to reflect its nature as a spinloop, possible future avenues for research if this loop plays hell with OS schedulers. Signed-off-by: Brian L. Troutwine <[email protected]> * Correct spelling of 'gague' to 'gauge' Signed-off-by: Brian L. Troutwine <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
As discussed in #6428 (comment) we should use
compare_exchange_weakwhere possible. It's worth taking a passes through the code to find opportunities for weaker ordering.The text was updated successfully, but these errors were encountered: