Re: [PATCH rcu 13/14] workqueue: Make queue_rcu_work() use call_rcu_flush()

[Joel Fernandes <joel@joelfernandes.org>]

> On Oct 31, 2022, at 9:21 AM, Uladzislau Rezki <urezki@gmail.com> wrote:
> 
> On Fri, Oct 28, 2022 at 09:23:47PM +0000, Joel Fernandes wrote:
>>> On Mon, Oct 24, 2022 at 09:48:19AM -0700, Paul E. McKenney wrote:
>>> On Mon, Oct 24, 2022 at 06:25:30PM +0200, Uladzislau Rezki wrote:
>>>>> 
>>>>> You guys might need to agree on the definition of "good" here.  Or maybe
>>>>> understand the differences in your respective platforms' definitions of
>>>>> "good".  ;-)
>>>>> 
>>>> Indeed. Bad is when once per-millisecond infinitely :) At least in such use
>>>> workload a can detect a power delta and power gain. Anyway, below is a new
>>>> trace where i do not use "flush" variant for the kvfree_rcu():
>>>> 
>>>> <snip>
>>>> 1. Home screen swipe:
>>>>         rcuop/0-15      [003] d..1  1792.767750: rcu_batch_start: rcu_preempt CBs=1003 bl=10
>>>>         rcuop/2-33      [002] d..1  1792.771717: rcu_batch_start: rcu_preempt CBs=934 bl=10
>>>>         rcuop/3-40      [001] d..1  1794.811816: rcu_batch_start: rcu_preempt CBs=1508 bl=11
>>>>         rcuop/1-26      [003] d..1  1797.116382: rcu_batch_start: rcu_preempt CBs=2127 bl=16
>>>>         rcuop/4-48      [001] d..1  1797.124422: rcu_batch_start: rcu_preempt CBs=95 bl=10
>>>>         rcuop/5-55      [002] d..1  1797.124731: rcu_batch_start: rcu_preempt CBs=143 bl=10
>>>>         rcuop/6-62      [005] d..1  1798.911719: rcu_batch_start: rcu_preempt CBs=132 bl=10
>>>>         rcuop/2-33      [002] d..1  1803.003966: rcu_batch_start: rcu_preempt CBs=3797 bl=29
>>>>         rcuop/0-15      [003] d..1  1803.004707: rcu_batch_start: rcu_preempt CBs=2969 bl=23
>>>> 2. App launches:
>>>>         rcuop/4-48      [005] d..1  1831.087612: rcu_batch_start: rcu_preempt CBs=6141 bl=47
>>>>         rcuop/7-69      [007] d..1  1831.095578: rcu_batch_start: rcu_preempt CBs=5464 bl=42
>>>>         rcuop/5-55      [004] d..1  1832.703571: rcu_batch_start: rcu_preempt CBs=8461 bl=66
>>>>         rcuop/0-15      [004] d..1  1833.731603: rcu_batch_start: rcu_preempt CBs=2548 bl=19
>>>>         rcuop/1-26      [006] d..1  1833.743691: rcu_batch_start: rcu_preempt CBs=2567 bl=20
>>>>         rcuop/2-33      [006] d..1  1833.744005: rcu_batch_start: rcu_preempt CBs=2359 bl=18
>>>>         rcuop/3-40      [006] d..1  1833.744286: rcu_batch_start: rcu_preempt CBs=3681 bl=28
>>>>         rcuop/4-48      [002] d..1  1838.079777: rcu_batch_start: rcu_preempt CBs=10444 bl=81
>>>>         rcuop/7-69      [001] d..1  1838.080375: rcu_batch_start: rcu_preempt CBs=12572 bl=98
>>>>           <...>-62      [002] d..1  1838.080646: rcu_batch_start: rcu_preempt CBs=14135 bl=110
>>>>         rcuop/6-62      [000] d..1  1838.087722: rcu_batch_start: rcu_preempt CBs=10839 bl=84

Please let us try to trim emails. That goes for me too.

>>>>           <...>-62      [003] d..1  1839.227022: rcu_batch_start: rcu_preempt CBs=1834 bl=14
>>>>           <...>-26      [001] d..1  1839.963315: rcu_batch_start: rcu_preempt CBs=5769 bl=45
>>>>         rcuop/2-33      [001] d..1  1839.966485: rcu_batch_start: rcu_preempt CBs=3789 bl=29
>>>>           <...>-40      [001] d..1  1839.966596: rcu_batch_start: rcu_preempt CBs=6425 bl=50
>>>>         rcuop/2-33      [005] d..1  1840.541272: rcu_batch_start: rcu_preempt CBs=825 bl=10
>>>>         rcuop/2-33      [005] d..1  1840.547724: rcu_batch_start: rcu_preempt CBs=44 bl=10
>>>>         rcuop/2-33      [005] d..1  1841.075759: rcu_batch_start: rcu_preempt CBs=516 bl=10
>>>>         rcuop/0-15      [002] d..1  1841.695716: rcu_batch_start: rcu_preempt CBs=6312 bl=49
>>>>         rcuop/0-15      [003] d..1  1841.709714: rcu_batch_start: rcu_preempt CBs=39 bl=10
>>>>         rcuop/5-55      [004] d..1  1843.112442: rcu_batch_start: rcu_preempt CBs=16007 bl=125
>>>>         rcuop/5-55      [004] d..1  1843.115444: rcu_batch_start: rcu_preempt CBs=7901 bl=61
>>>>         rcuop/6-62      [001] dn.1  1843.123983: rcu_batch_start: rcu_preempt CBs=8427 bl=65
>>>>         rcuop/6-62      [006] d..1  1843.412383: rcu_batch_start: rcu_preempt CBs=981 bl=10
>>>>         rcuop/0-15      [003] d..1  1844.659812: rcu_batch_start: rcu_preempt CBs=1851 bl=14
>>>>         rcuop/0-15      [003] d..1  1844.667790: rcu_batch_start: rcu_preempt CBs=135 bl=10
>>>> <snip>
>>>> 
>>>> it is much more better. But. As i wrote earlier there is a patch that i have submitted
>>>> some time ago improving kvfree_rcu() batching:
>>>> 
>>>> <snip>
>>>> commit 51824b780b719c53113dc39e027fbf670dc66028
>>>> Author: Uladzislau Rezki (Sony) <urezki@gmail.com>
>>>> Date:   Thu Jun 30 18:33:35 2022 +0200
>>>> 
>>>>    rcu/kvfree: Update KFREE_DRAIN_JIFFIES interval
>>>> 
>>>>    Currently the monitor work is scheduled with a fixed interval of HZ/20,
>>>>    which is roughly 50 milliseconds. The drawback of this approach is
>>>>    low utilization of the 512 page slots in scenarios with infrequence
>>>>    kvfree_rcu() calls.  For example on an Android system:
>>>> <snip>
>>>> 
>>>> The trace that i posted was taken without it.
>>> 
>>> And if I am not getting too confused, that patch is now in mainline.
>>> So it does make sense to rely on it, then.  ;-)
>> 
>> Vlad's patch to change the KFREE_DRAIN_JIFFIES to 5 seconds seems reasonable
>> to me. However, can we unify KFREE_DRAIN_JIFFIES and LAZY_FLUSH_JIFFIES ?
>> 
> This is very good.
> 
> Below is a plot that i have taken during one use-case. It is about three
> apps usage in parallel. It was done by running "monkey" test:
> 
> wget ftp://vps418301.ovh.net/incoming/monkey_3_apps_slab_usage_5_minutes.png
> 
> i set up three apps as usage scenario: Google Chrome, YoTube and Camera.
> I logged the Slab metric from the /proc/meminfo. Sampling rate is 0.1 second.
> 
> Please have a look at results. It reflects what i am saying. non-flush
> kvfree RCU variant makes a memory usage higher. What is not acceptable
> for our mobile devices and workloads.

Thanks I’ll take a closer look at the data (currently commuting) but here’s a quick reply:

I am curious that with the 5 second timer, you are delaying RCU anyway. Are you saying that, adding another 10 on top (due to lazyfying) seems to be causing issues? I find it hard to believe that you cannot give the shrinker enough work within 5 seconds, such that it also triggers the issues you’re seeing. However the workload and data speaks.

>> One at 5 and other at 10 seems odd, especially because the former seems to
>> negate the effects of the latter and anyone tweaking that in the future (say
>> via new command line options) should probably tweak them together to increase
>> batching.
>> 
> Well. Convert 5 seconds to 10? What will it solve for you? We can do it
> and from a kvfree_rcu() perspective nothing really is changed.

True. In fact with my last patch, I see almost never even the need to go to RCU. However my point with unification is just to keep it simple for user for 2 knobs that do the same thing. Granted this is a compiler knob but that might change in the future. We already have enough knobs in RCU and as you guys know, I’m a fan of not letting the user mess things up too much.

>> Testing shows significantly better batching with Vlad's updates, however I am
>> wondering why the rcu_callback fires in pairs like that from separate
>> kworkers:
>> 
>>     kworker/6:1-157     [006] d..1   288.861610: rcu_callback: rcu_preempt  rhp=0000000079b895f9 func=rcu_work_rcufn 1214
>>     kworker/4:2-158     [004]  d..1   288.861612: rcu_callback: rcu_preempt rhp=00000000d83fcc90 func=rcu_work_rcufn 798
>> 
>> I wonder if the queued kwork is executing twice accidentally, or something.
>> 
> Because a kfree_rcu_cpu() is a per-cpu thing.

Right, got it.

>> This kernel does have the additional trace patch below, fyi.
>> 
>> Another thought I have is, if we can just keep the kvfree_rcu() mapped to
>> call_rcu() via a config option say CONFIG_RCU_LAZY_KFREE, or something.
>> 
> I am not sure you need it, really. If you wake-up "rcuop" or whatever
> with 0.5 second interval or with 5 seconds interval you will not notice
> anything in terms of power between both.

Yes, you are right. This is not needed considering the improvements you recently made.

Cheers,

 - Joel 


> 
> --
> Uladzislau Rezki