On Tue, 2008-06-24 at 07:15 -0600, Gregory Haskins wrote: > >>> On Tue, Jun 24, 2008 at 6:13 AM, in message <1214302405.4351.21.camel@twins>, > Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > On Mon, 2008-06-23 at 17:04 -0600, Gregory Haskins wrote: > >> We do find_busiest_groups() et. al. without locks held for normal balancing, > >> so lets do it for newidle as well. It will allow other cpus to make > >> forward progress (against our RQ) while we try to balance and allow > >> some interrupts to occur. > > > > Is running f_b_g really that expensive? > > According to our oprofile data, yes. I speculate that it works out that way because most newidle > attempts result in "no imbalance". But we were spending ~60%+ time in find_busiest_groups() > because of all the heavy-context switching that goes on in PREEMPT_RT. So while f_b_g() is > probably cheaper than double-lock/move_tasks(), the ratio of occurrence is off the charts in > comparison. Prior to this patch, those occurrences were preempt-disabled/irq-disabled/rq->lock critical > sections. > > So while it is not clear if f_b_g() is the actual cost, it is a convenient (and legal, afaict) place to > deterministically reduce the rq->lock scope. Additionally, doing so measurably helps > performance, so I think its a win. Without this patch you have to hope the double_lock releases > this_rq, and even so were not checking for the NEEDS_RESCHED. See, having had this information in the changelog to begin with would have helped ;-) -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
