On Tue, 2014-07-15 at 14:59 +0200, Thomas Gleixner wrote: > On Tue, 15 Jul 2014, Peter Zijlstra wrote: > > > On Tue, Jul 15, 2014 at 11:50:45AM +0200, Peter Zijlstra wrote: > > > So you already have an idle notifier (which is x86 only, we should fix > > > that I suppose), and you then double check there really isn't anything > > > else running. > > > > Note that we've already done a large part of the expense of going idle > > by the time we call that idle notifier -- in specific, we've > > reprogrammed the clock to stop the tick. > > > > Its really wasteful to then generate work again, which means we have to > > again reprogram the clock etc. > > Doing anything which is not related to idle itself in the idle > notifier is just plain wrong. > > If that stuff wants to utilize idle slots, we really need to come up > with a generic and general solution. Otherwise we'll grow those warts > all over the architecture space, with slightly different ways of > wreckaging the world an some more. > > This whole attidute of people thinking that they need their own > specialized scheduling around the real scheduler is a PITA. All this > stuff is just damanging any sensible approach of power saving, load > balancing, etc. Not to mention that we're already too rotund... pipe-test scheduling cross core, ie ~0 work, ~pure full fastpath. All kernels with same (obese) distro config, with drivers reduced to what my boxen need. Squint a little, there is some jitter. These kernels are all adjusted to eliminate various regressions that would otherwise skew results up to and including _very_ badly. See "virgin", the numbers are much more useful without that particular skew methinks :) 3.0.101-default 3.753363 usecs/loop -- avg 3.770737 530.4 KHz 1.000 3.1.10-default 3.723843 usecs/loop -- avg 3.716058 538.2 KHz 1.014 3.2.51-default 3.728060 usecs/loop -- avg 3.710372 539.0 KHz 1.016 3.3.8-default 3.906174 usecs/loop -- avg 3.900399 512.8 KHz .966 3.4.97-default 3.864158 usecs/loop -- avg 3.865281 517.4 KHz .975 3.5.7-default 3.967481 usecs/loop -- avg 3.962757 504.7 KHz .951 3.6.11-default 3.851186 usecs/loop -- avg 3.845321 520.1 KHz .980 3.7.10-default 3.777869 usecs/loop -- avg 3.776913 529.5 KHz .998 3.8.13-default 4.049927 usecs/loop -- avg 4.041905 494.8 KHz .932 3.9.11-default 3.973046 usecs/loop -- avg 3.974208 503.2 KHz .948 3.10.27-default 4.189598 usecs/loop -- avg 4.189298 477.4 KHz .900 3.11.10-default 4.293870 usecs/loop -- avg 4.297979 465.3 KHz .877 3.12.24-default 4.321570 usecs/loop -- avg 4.321961 462.8 KHz .872 3.13.11-default 4.137845 usecs/loop -- avg 4.134863 483.7 KHz .911 3.14.10-default 4.145348 usecs/loop -- avg 4.139987 483.1 KHz .910 1.000 3.15.4-default 4.355594 usecs/loop -- avg 4.351961 459.6 KHz .866 .951 1.000 3.16.0-default 4.537279 usecs/loop -- avg 4.543532 440.2 KHz .829 .911 .957 3.16.0-virgin 6.377331 usecs/loop -- avg 6.352794 314.8 KHz 0.sob my local config, group sched, namespaces etc disabled 3.0.101-smp 3.692377 usecs/loop -- avg 3.690774 541.9 KHz 1.000 3.1.10-smp 3.573832 usecs/loop -- avg 3.563269 561.3 KHz 1.035 3.2.51-smp 3.632690 usecs/loop -- avg 3.628220 551.2 KHz 1.017 3.3.8-smp 3.801838 usecs/loop -- avg 3.803441 525.8 KHz .970 3.4.97-smp 3.836087 usecs/loop -- avg 3.843501 520.4 KHz .960 3.5.7-smp 3.646927 usecs/loop -- avg 3.646288 548.5 KHz 1.012 3.6.11-smp 3.674402 usecs/loop -- avg 3.680929 543.3 KHz 1.002 3.7.10-smp 3.644274 usecs/loop -- avg 3.644566 548.8 KHz 1.012 3.8.13-smp 3.678164 usecs/loop -- avg 3.675524 544.1 KHz 1.004 3.9.11-smp 3.834943 usecs/loop -- avg 3.845852 520.0 KHz .959 3.10.27-smp 3.651881 usecs/loop -- avg 3.634515 550.3 KHz 1.015 3.11.10-smp 3.716159 usecs/loop -- avg 3.720603 537.5 KHz .991 3.12.24-smp 3.862634 usecs/loop -- avg 3.872252 516.5 KHz .953 3.13.11-smp 3.803254 usecs/loop -- avg 3.802553 526.0 KHz .970 3.14.10-smp 4.010009 usecs/loop -- avg 4.009019 498.9 KHz .920 3.15.4-smp 3.882398 usecs/loop -- avg 3.884095 514.9 KHz .950 3.16.0-master 4.061003 usecs/loop -- avg 4.058244 492.8 KHz .909 echo 0 > sched_wakeup_granularity_ns, taskset -c 3 pipe-test 1 (shortest path) 3.0.101-default 3.352267 usecs/loop -- avg 3.352434 596.6 KHz 1.000 3.16.0-default 3.596559 usecs/loop -- avg 3.594023 556.5 KHz .932 3.0.101-smp 3.089251 usecs/loop -- avg 3.089556 647.3 KHz 1.000 3.16.0-master 3.254721 usecs/loop -- avg 3.251534 615.1 KHz .950 sched+idle is becoming more of a not-so-fastpath. Pure sched is not as bad, but still, we're getting fat. netperf TCP_RR trans/sec (unbound) 3.0.101-default 91360.56 1.000 3.16.0-default 72523.30 .793 3.0.101-smp 92166.23 1.000 3.16.0-master 81235.30 .881 echo 0 > sched_wakeup_granularity_ns, bound to cpu3 3.0.101-smp 94289.95 1.000 3.16.0-master 81219.02 .861 Leanest meanest kernel ever to run on this box (2.6.22 + cfs-2.6.25 etc) did that bound TCP_RR at ~114k IIRC. My userspace became too new to boot that kernel without a squabble, but I think I recall correctly. -Mike -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
