linux-kernel - Re: [REGRESSION] 5-10% increase in IO latencies with nohz balance patch

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

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <YeBZ7qNjPsonEYZz@carbon.dhcp.thefacebook.com>
Date:   Thu, 13 Jan 2022 08:57:18 -0800
From:   Roman Gushchin <guro@...com>
To:     Valentin Schneider <valentin.schneider@....com>
CC:     Josef Bacik <josef@...icpanda.com>,
        Thorsten Leemhuis <regressions@...mhuis.info>,
        <peterz@...radead.org>, <vincent.guittot@...aro.org>,
        <torvalds@...ux-foundation.org>, <linux-kernel@...r.kernel.org>,
        <linux-btrfs@...r.kernel.org>, <clm@...com>
Subject: Re: [REGRESSION] 5-10% increase in IO latencies with nohz balance
 patch

On Thu, Jan 13, 2022 at 04:41:57PM +0000, Valentin Schneider wrote:
> On 03/01/22 11:16, Josef Bacik wrote:
> > On Wed, Dec 22, 2021 at 04:07:35PM +0000, Valentin Schneider wrote:
> >> 
> >> Hi,
> >> 
> >> On 22/12/21 13:42, Thorsten Leemhuis wrote:
> >> > What's the status here? Just wondering, because there hasn't been any
> >> > activity in this thread since 11 days and the festive season is upon us.
> >> >
> >> > Was the discussion moved elsewhere? Or is this still a mystery? And if
> >> > it is: how bad is it, does it need to be fixed before Linus releases 5.16?
> >> >
> >> 
> >> I got to the end of bisect #3 yesterday, the incriminated commit doesn't
> >> seem to make much sense but I've just re-tested it and there is a clear
> >> regression between that commit and its parent (unlike bisect #1 and #2):
> >> 
> >> 2127d22509aec3a83dffb2a3c736df7ba747a7ce mm, slub: fix two bugs in slab_debug_trace_open()
> >> write_clat_ns_p99     195395.92     199638.20      4797.01    2.17%
> >> write_iops             17305.79      17188.24       250.66   -0.68%
> >> 
> >> write_clat_ns_p99     195543.84     199996.70      5122.88    2.28%
> >> write_iops             17300.61      17241.86       251.56   -0.34%
> >> 
> >> write_clat_ns_p99     195543.84     200724.48      5122.88    2.65%
> >> write_iops             17300.61      17246.63       251.56   -0.31%
> >> 
> >> write_clat_ns_p99     195543.84     200445.41      5122.88    2.51%
> >> write_iops             17300.61      17215.47       251.56   -0.49%
> >> 
> >> 6d2aec9e123bb9c49cb5c7fc654f25f81e688e8c mm/mempolicy: do not allow illegal MPOL_F_NUMA_BALANCING | MPOL_LOCAL in mbind() 
> >> write_clat_ns_p99     195395.92     197942.30      4797.01    1.30%
> >> write_iops             17305.79      17246.56       250.66   -0.34%
> >> 
> >> write_clat_ns_p99     195543.84     196183.92      5122.88    0.33%
> >> write_iops             17300.61      17310.33       251.56    0.06%
> >> 
> >> write_clat_ns_p99     195543.84     196990.71      5122.88    0.74%
> >> write_iops             17300.61      17346.32       251.56    0.26%
> >> 
> >> write_clat_ns_p99     195543.84     196362.24      5122.88    0.42%
> >> write_iops             17300.61      17315.71       251.56    0.09%
> >> 
> >> It's pure debug stuff and AFAICT is a correct fix...
> >> @Josef, could you test that on your side?
> >
> > Sorry, holidays and all that.  I see 0 difference between the two commits, and
> > no regression from baseline.  It'll take me a few days to recover from the
> > holidays, but I'll put some more effort into actively debugging wtf is going on
> > here on my side since we're all having trouble pinning down what's going
> > on.
> 
> Humph, that's unfortunate... I just came back from my holidays, so I'll be
> untangling my inbox for the next few days. Do keep us posted!

I'm trying to bisect it independently and make sense of it too, thanks to Josef
for providing me a test setup. From the very first data I've got yesterday,
the only thing I can say the data is very noisy and I'm not totally convinced
that the regression is coming from the patch which was blamed initially.

I hope to make more progress today/tomorrow, will keep you updated.

Thanks!