Re: [PATCH] sched/fair: Again ignore percpu threads for imbalance pulls

From: Yihao Wu
Date: Tue Jan 18 2022 - 03:11:23 EST


On 2022/1/18 1:16 am, Valentin Schneider wrote:
On 17/01/22 22:50, Yihao Wu wrote:
Thanks a lot for the help, Valentin and Peter!

On 2021/12/17 2:26am, Valentin Schneider wrote:
On 11/12/21 17:48, Yihao Wu wrote:
commit 2f5f4cce496e ("sched/fair: Ignore percpu threads for imbalance
pulls") was meant to fix a performance issue, when load balance tries to
migrate pinned kernel threads at MC domain level. This was destined to
fail.

After it fails, it further makes wakeup balance at NUMA domain level
messed up. The most severe case that I noticed and frequently occurs:
|sum_nr_running(node1) - sum_nr_running(node2)| > 100


Wakeup balance (aka find_idlest_cpu()) is different from periodic load
balance (aka load_balance()) and doesn't use can_migrate_task(), so the
incriminated commit shouldn't have impacted it (at least not in obvious
ways...). Do you have any more details on that issue

The original bugfix concerns only about load balance. While I found wake
up balance is impacted too, after I observed regression in lmbench3 test
suite. This is how it's impacted:

- Periodic load balance
- kthread_is_per_cpu? No
- env->flags |= LBF_SOME_PINNED
- sd_parent..imbalance being set to 1 because of LBF_SOME_PINNED

So far exactly the same as what Chandrasekhar describes in 2f5f4cce496e.
Then imbalance connects periodic and wakeup balance.

- Wakeup balance(find_idlest_group)
- update_sg_wakeup_stats classifies local_sgs as group_imbalanced
- find_idlest_group chooses another NUMA node

wakeup balance keeps doing this until another NUMA node becomes so busy.
And another periodic load balance just shifts it around, makeing the
previously overloaded node completely idle now.


Oooh, right, I came to the same conclusion when I got that stress-ng
regression report back then:

https://lore.kernel.org/all/871rajkfkn.mognet@xxxxxxx/


Shocked! I wasted weeks to locate almost the same regression. Why on
earth haven't I read your discussion of half a year ago?

I pretty much gave up on that as the regression we caused by removing an
obscure/accidental balance which I couldn't properly codify. I can give it

Strange, the regression reported to me says differently from yours.

4.19.91 before_2f5f4 after_2f5f4
my_report good bad bad
your_report N/A good bad

your_report says 2f5f4 introduces new regression. While
my_report says 2f5f4 fails and leaves the old regression be ...

Maybe that's the reason why you give up on fixing it, yet I came to make can_migrate_task cover more cases (kernel_thread).


another shot, but AFAICT that only affects fork/exec heavy workloads (that
-13% was on something doing almost only forks) which is an odd case to
support.

Yes. They're indeed quite odd workloads.
- Apps with massive shortlived threads better change runtime model, or
use a thread pool.
- Massive different apps on the same machine are even odder.

But I guess this problem affects normal workloads too, more or less but
not significantly. Hard to tell exactly how much influence it has.

(Thanks to the great schedviz tool, I observed that all workloads as a
whole, is migrated between the two NUMA nodes in a ping-pong pattern,
and with a period around 3ms)

The reason wake up balance suffers more is, in fork+exit test case,
wakeup balance happens with much higher frequency. It exists in real
world applications too I believe.


However the original bugfix failed, because it covers only case 1) below.
1) Created by create_kthread
2) Created by kernel_thread
No kthread is assigned to task_struct in case 2 (Please refer to comments
in free_kthread_struct) so it simply won't work.

The easist way to cover both cases is to check nr_cpus_allowed, just as
discussed in the mailing list of the v1 version of the original fix.

* lmbench3.lat_proc -P 104 fork (2 NUMA, and 26 cores, 2 threads)


Reasoning about "proper" pcpu kthreads was simpler since they are static,
see 3a7956e25e1d ("kthread: Fix PF_KTHREAD vs to_kthread() race")

Get it. Thanks.

w/out patch w/ patch
fork+exit latency 1660 ms 1520 ms ( 8.4%)

Fixes: 2f5f4cce496e ("sched/fair: Ignore percpu threads for imbalance pulls")
Signed-off-by: Yihao Wu <wuyihao@xxxxxxxxxxxxxxxxx>
---
kernel/kthread.c | 6 +-----
1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/kernel/kthread.c b/kernel/kthread.c
index 4a4d7092a2d8..cb05d3ff2de4 100644
--- a/kernel/kthread.c
+++ b/kernel/kthread.c
@@ -543,11 +543,7 @@ void kthread_set_per_cpu(struct task_struct *k, int cpu)

bool kthread_is_per_cpu(struct task_struct *p)
{
- struct kthread *kthread = __to_kthread(p);
- if (!kthread)
- return false;
-
- return test_bit(KTHREAD_IS_PER_CPU, &kthread->flags);
+ return (p->flags & PF_KTHREAD) && p->nr_cpus_allowed == 1;
}

As Peter said, this is going to cause issues. If you look at
kthread_set_per_cpu(), we also store a CPU value which we expect to be
valid when kthread_is_per_cpu(), which that change is breaking.

AIUI what you want to patch is the actual usage in can_migrate_task()


Get it. Some may want a consistent view of kthread_is_per_cpu,
kthread->cpu, and KTHREAD_IS_PER_CPU.

Are you suggesting to patch only can_migrate_task to check
nr_cpus_allowed?

Yes


Okay, I'll post a v2. And see if Peter likes it.

Wouldn't it be confusing if it uses an alternative way
to tell if p is a per-cpu kthread?


Well then it wouldn't catch just per-CPU kthreads, but rather any pinned
task (kernel or otherwise). But then you have to check/test if that's a
sane thing to :)


Sounds like pain... and not an option :-D

Thanks,
Yihao Wu

I haven't a better solution though. :(


Thanks,
Yihao Wu


/**
--
2.32.0.604.gb1f3e1269