Messages in this thread | | | Date | Wed, 9 Apr 2008 22:10:10 +0400 | From | Alexey Dobriyan <> | Subject | Re: [patch 09/17] Add all cpus option to stop machine run |
| |
Please, stop Cc'ing me on this.
I continue to think that all this so-called "immediate" infrastructure is an absolute overkill and declared D-cache line savings will _never_ matter and will never be measured on real life workloads. Right now you claim 1 (one) cacheline saving in schedule() which can be trivially moved under CONFIG_PROFILING umbrella. On the other hand is more ugly assembler code. Sorry, people add infra with much better good/bad ratios.
Also, my gut feeling of a guy who is also on receiveing end of bugreports at SWsoft that line with .text games was crossed by SMP alternatives and fooprobes. The rest won't matter because programs like firefox will fuck up L1, L2 caches in one blow just by showing animation.
We don't need bugs when immediate update screwed up fighting for cacheline.
And bugs when screwup happened out of the window of "Code:" printout, so you won't have a chance to compare relevant vmlinux .text and printout.
And bugs when CPU executed bullshit and silently rebooted.
And so on.
Alexey, more and more liking OpenBSD where such games won't even hit mailing lists
| |