Messages in this thread | | | Date | Tue, 19 Aug 2008 01:04:28 -0700 | From | Jeremy Fitzhardinge <> | Subject | Re: [PATCH 0 of 9] x86/smp function calls: convert x86 tlb flushes to use function calls [POST 2] |
| |
Andi Kleen wrote: >> Don't pay attention to that number. It's only the extra latency of a HT >> context->HT context function call + tlb flush. Which means 1) the tlb >> is shared anyway, so the extra flush is redundant, 2) they're not really >> concurrent, 3) it's going down the single-cpu call, rather than the >> multi-cpu one, 4) it's only measuring the latency for a particular tlb >> flush, and doesn't take into account any throughput improvements the >> extra queueing may add. >> > > A lot of flushes are synchronous. See the rest of my email that > you snipped.
Yep, sure, but I just wanted to be clear that the 35% number is almost useless in isolation, and the first step is to do proper measurements.
J
| |