| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <20131220111926.GA11605@e103034-lin>
Date: Fri, 20 Dec 2013 11:19:26 +0000
From: Morten Rasmussen <morten.rasmussen@....com>
To: Alex Shi <alex.shi@...aro.org>
Cc: Peter Zijlstra <peterz@...radead.org>,
"mingo@...hat.com" <mingo@...hat.com>,
"vincent.guittot@...aro.org" <vincent.guittot@...aro.org>,
"daniel.lezcano@...aro.org" <daniel.lezcano@...aro.org>,
"fweisbec@...il.com" <fweisbec@...il.com>,
"linux@....linux.org.uk" <linux@....linux.org.uk>,
"tony.luck@...el.com" <tony.luck@...el.com>,
"fenghua.yu@...el.com" <fenghua.yu@...el.com>,
"tglx@...utronix.de" <tglx@...utronix.de>,
"akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
"arjan@...ux.intel.com" <arjan@...ux.intel.com>,
"pjt@...gle.com" <pjt@...gle.com>,
"fengguang.wu@...el.com" <fengguang.wu@...el.com>,
"james.hogan@...tec.com" <james.hogan@...tec.com>,
"jason.low2@...com" <jason.low2@...com>,
"gregkh@...uxfoundation.org" <gregkh@...uxfoundation.org>,
"hanjun.guo@...aro.org" <hanjun.guo@...aro.org>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 4/4] sched: bias to target cpu load to reduce task moving
On Thu, Dec 19, 2013 at 01:34:08PM +0000, Alex Shi wrote:
> On 12/17/2013 11:38 PM, Peter Zijlstra wrote:
> > On Tue, Dec 17, 2013 at 02:10:12PM +0000, Morten Rasmussen wrote:
> >>> @@ -4135,7 +4141,7 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
> >>> if (local_group)
> >>> load = source_load(i);
> >>> else
> >>> - load = target_load(i);
> >>> + load = target_load(i, sd->imbalance_pct);
> >>
> >> Don't you apply imbalance_pct twice here? Later on in
> >> find_idlest_group() you have:
> >>
> >> if (!idlest || 100*this_load < imbalance*min_load)
> >> return NULL;
> >>
> >> where min_load comes from target_load().
> >
> > Yes! exactly! this doesn't make any sense.
>
> Thanks a lot for review and comments!
>
> I changed the patch to following shape. and push it under Fengguang's testing
> system monitor. Any testing are appreciated!
>
> BTW, Seems lots of changes in scheduler come from kinds of scenarios/benchmarks
> experience. But I still like to take any theoretical comments/suggestions.
>
> --
> Thanks
> Alex
>
> ===
>
> From 5cd67d975001edafe2ee820e0be5d86881a23bd6 Mon Sep 17 00:00:00 2001
> From: Alex Shi <alex.shi@...aro.org>
> Date: Sat, 23 Nov 2013 23:18:09 +0800
> Subject: [PATCH 4/4] sched: bias to target cpu load to reduce task moving
>
> Task migration happens when target just a bit less then source cpu load.
> To reduce such situation happens, aggravate the target cpu load with
> sd->imbalance_pct/100 in wake_affine.
>
> In find_idlest/busiest_group, change the aggravate to local cpu only
> from old group aggravation.
>
> on my pandaboard ES.
>
> latest kernel 527d1511310a89 + whole patchset
> hackbench -T -g 10 -f 40
> 23.25" 21.99"
> 23.16" 21.20"
> 24.24" 21.89"
> hackbench -p -g 10 -f 40
> 26.52" 21.46"
> 23.89" 22.96"
> 25.65" 22.73"
> hackbench -P -g 10 -f 40
> 20.14" 19.72"
> 19.96" 19.10"
> 21.76" 20.03"
>
> Signed-off-by: Alex Shi <alex.shi@...aro.org>
> ---
> kernel/sched/fair.c | 35 ++++++++++++++++-------------------
> 1 file changed, 16 insertions(+), 19 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index bccdd89..3623ba4 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -978,7 +978,7 @@ static inline unsigned long group_weight(struct task_struct *p, int nid)
>
> static unsigned long weighted_cpuload(const int cpu);
> static unsigned long source_load(int cpu);
> -static unsigned long target_load(int cpu);
> +static unsigned long target_load(int cpu, int imbalance_pct);
> static unsigned long power_of(int cpu);
> static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
>
> @@ -3809,11 +3809,17 @@ static unsigned long source_load(int cpu)
> * Return a high guess at the load of a migration-target cpu weighted
> * according to the scheduling class and "nice" value.
> */
> -static unsigned long target_load(int cpu)
> +static unsigned long target_load(int cpu, int imbalance_pct)
> {
> struct rq *rq = cpu_rq(cpu);
> unsigned long total = weighted_cpuload(cpu);
>
> + /*
> + * without cpu_load decay, in most of time cpu_load is same as total
> + * so we need to make target a bit heavier to reduce task migration
> + */
> + total = total * imbalance_pct / 100;
> +
> if (!sched_feat(LB_BIAS))
> return total;
>
> @@ -4033,7 +4039,7 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
> this_cpu = smp_processor_id();
> prev_cpu = task_cpu(p);
> load = source_load(prev_cpu);
> - this_load = target_load(this_cpu);
> + this_load = target_load(this_cpu, 100);
>
> /*
> * If sync wakeup then subtract the (maximum possible)
> @@ -4089,7 +4095,7 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
>
> if (balanced ||
> (this_load <= load &&
> - this_load + target_load(prev_cpu) <= tl_per_task)) {
> + this_load + target_load(prev_cpu, 100) <= tl_per_task)) {
> /*
> * This domain has SD_WAKE_AFFINE and
> * p is cache cold in this domain, and
> @@ -4112,7 +4118,6 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
> {
> struct sched_group *idlest = NULL, *group = sd->groups;
> unsigned long min_load = ULONG_MAX, this_load = 0;
> - int imbalance = 100 + (sd->imbalance_pct-100)/2;
>
> do {
> unsigned long load, avg_load;
> @@ -4132,10 +4137,10 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
>
> for_each_cpu(i, sched_group_cpus(group)) {
> /* Bias balancing toward cpus of our domain */
> - if (local_group)
> + if (i == this_cpu)
What is the motivation for changing the local_group load calculation?
Now the load contributions of all cpus in the local group, except
this_cpu, will contribute more as their contribution (this_load) is
determined using target_load() instead.
If I'm not mistaken, that will lead to more frequent load balancing as
the local_group bias has been reduced. That is the opposite of your
intentions based on your comment in target_load().
> load = source_load(i);
> else
> - load = target_load(i);
> + load = target_load(i, sd->imbalance_pct);
You scale by sd->imbalance_pct instead of 100+(sd->imbalance_pct-100)/2
that you removed above. sd->imbalance_pct may have been arbitrarily
chosen in the past, but changing it may affect behavior.
>
> avg_load += load;
> }
> @@ -4151,7 +4156,7 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
> }
> } while (group = group->next, group != sd->groups);
>
> - if (!idlest || 100*this_load < imbalance*min_load)
> + if (!idlest || this_load < min_load)
> return NULL;
> return idlest;
> }
> @@ -5476,9 +5481,9 @@ static inline void update_sg_lb_stats(struct lb_env *env,
>
> nr_running = rq->nr_running;
>
> - /* Bias balancing toward cpus of our domain */
> - if (local_group)
> - load = target_load(i);
> + /* Bias balancing toward dst cpu */
> + if (env->dst_cpu == i)
> + load = target_load(i, env->sd->imbalance_pct);
Here you do the same group load bias change as above.
> else
> load = source_load(i);
>
> @@ -5918,14 +5923,6 @@ static struct sched_group *find_busiest_group(struct lb_env *env)
> if ((local->idle_cpus < busiest->idle_cpus) &&
> busiest->sum_nr_running <= busiest->group_weight)
> goto out_balanced;
> - } else {
> - /*
> - * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
> - * imbalance_pct to be conservative.
> - */
> - if (100 * busiest->avg_load <=
> - env->sd->imbalance_pct * local->avg_load)
> - goto out_balanced;
> }
>
> force_balance:
As said my previous replies to this series, I think this problem should
be solved by fixing the cause of the problem, that is the cpu_load
calculation, instead of biasing the cpu_load where-ever it is used to
hide the problem.
Doing a bit of git archaeology reveals that the cpu_load code goes back
to 7897986bad8f6cd50d6149345aca7f6480f49464 and that in the original
patch x86 was using *_idx > 0 for all sched_domain levels except SMT. In
my opinion that made logical sense. If we are about to change to *_idx=0
we are removing the main idea behind that code and there needs to be a
new one. Otherwise, cpu_load doesn't make sense.
Morten
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists