Messages in this thread | | | Subject | Re: [PATCH] cpufreq: intel_pstate: Handle no_turbo in frequency invariance | From | Giovanni Gherdovich <> | Date | Fri, 08 Apr 2022 16:22:28 +0200 |
| |
On Fri, 2022-04-08 at 07:42 +0800, Chen Yu wrote: > Problem statement: > Once the user has disabled turbo frequency by > echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo, > the cfs_rq's util_avg becomes quite small when compared with > CPU capacity. > > Step to reproduce: > > echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo > > ./x86_cpuload --count 1 --start 3 --timeout 100 --busy 99 > would launch 1 thread and bind it to CPU3, lasting for 100 seconds, > with a CPU utilization of 99%. [1] > > top result: > %Cpu3 : 98.4 us, 0.0 sy, 0.0 ni, 1.6 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st > > check util_avg: > cat /sys/kernel/debug/sched/debug | grep "cfs_rq\[3\]" -A 20 | grep util_avg > .util_avg : 611 > > So the util_avg/cpu capacity is 611/1024, which is much smaller than > 98.4% shown in the top result. > > This might impact some logic in the scheduler. For example, group_is_overloaded() > would compare the group_capacity and group_util in the sched group, to > check if this sched group is overloaded or not. With this gap, even > when there is a nearly 100% workload, the sched group will not be regarded > as overloaded. Besides group_is_overloaded(), there are also other victims. > There is a ongoing work that aims to optimize the task wakeup in a LLC domain. > The main idea is to stop searching idle CPUs if the sched domain is overloaded[2]. > This proposal also relies on the util_avg/CPU capacity to decide whether the LLC > domain is overloaded. > > Analysis: > CPU frequency invariance has caused this difference. In summary, > the util_sum of cfs rq would decay quite fast when the CPU is in > idle, when the CPU frequency invariance is enabled. > > The detail is as followed: > > As depicted in update_rq_clock_pelt(), when the frequency invariance > is enabled, there would be two clock variables on each rq, clock_task > and clock_pelt: > > The clock_pelt scales the time to reflect the effective amount of > computation done during the running delta time but then syncs back to > clock_task when rq is idle. > > absolute time | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16 > @ max frequency ------******---------------******--------------- > @ half frequency ------************---------************--------- > clock pelt | 1| 2| 3| 4| 7| 8| 9| 10| 11|14|15|16 > > The fast decay of util_sum during idle is due to: > 1. rq->clock_pelt is always behind rq->clock_task > 2. rq->last_update is updated to rq->clock_pelt' after invoking ___update_load_sum() > 3. Then the CPU becomes idle, the rq->clock_pelt' would be suddenly increased > a lot to rq->clock_task > 4. Enters ___update_load_sum() again, the idle period is calculated by > rq->clock_task - rq->last_update, AKA, rq->clock_task - rq->clock_pelt'. > The lower the CPU frequency is, the larger the delta = > rq->clock_task - rq->clock_pelt' will be. Since the idle period will be > used to decay the util_sum only, the util_sum drops significantly during > idle period. > > Proposal: > This symptom is not only caused by disabling turbo frequency, but it > would also appear if the user limits the max frequency at runtime. Because > if the frequency is always lower than the max frequency, > CPU frequency invariance would decay the util_sum quite fast during idle. > > As some end users would disable turbo after boot up, this patch aims to > present this symptom and deals with turbo scenarios for now. It might > be ideal if CPU frequency invariance is aware of the max CPU frequency > (user specified) at runtime in the future. > > [Previous patch seems to be lost on LKML, this is a resend, sorry for any > inconvenience] > > Link: https://github.com/yu-chen-surf/x86_cpuload.git #1 > Link: https://lore.kernel.org/lkml/20220310005228.11737-1-yu.c.chen@intel.com/ #2 > Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Reviewed-by: Giovanni Gherdovich <ggherdovich@suse.cz>
You're right, when turbo is disabled, the frequency invariance code needs to know about it; it calculates freq_curr/freq_max and thinks that freq_max is some turbo level. For example commit 918229cdd5ab ("x86/intel_pstate: Handle runtime turbo disablement/enablement in frequency invariance") takes care of this when global.turbo_disabled changes, but before your patch nothing checks if the user disabled turbo from sysfs. Thanks for the fix!
Giovanni
> --- > drivers/cpufreq/intel_pstate.c | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c > index 846bb3a78788..2216b24b6f84 100644 > --- a/drivers/cpufreq/intel_pstate.c > +++ b/drivers/cpufreq/intel_pstate.c > @@ -1322,6 +1322,7 @@ static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b, > mutex_unlock(&intel_pstate_limits_lock); > > intel_pstate_update_policies(); > + arch_set_max_freq_ratio(global.no_turbo); > > mutex_unlock(&intel_pstate_driver_lock); >
| |