On Thu Jul 11 16:27:08 2024 +0000, Tatsuyuki Ishi wrote:
I've already resolved this offline with Gofman because having a productive discussion with a confrontational, overconfident and unprofessional "academic" is just impossible. Just leave it to Huw for review, the description already says it all.
@ishitatsuyuki I'll try to explain to you, you suggest using `gettimeofday`, which was abandoned in the commit you want to revert. But calling `gettimeofday` may take significantly longer and on older kernels the call may not be wrapped in a `vDSO (virtual dynamic shared object)`. Also, the current behavior is to use `CLOCK_REALTIME_COARSE` in `NtDelayExecution`, I think it’s a bad idea because we can get time changes in `CLOCK_REALTIME_COARSE`, from here our timer can be offset in a non-predictable way.
``` This will only affect users running with HZ=1000. On an i7-8700 CPU @ 3.20GHz it cuts the call cost from ~30ns to ~12ns. ```