Hi Paolo et al.
I have a strong suspect that something is going wrong when the
underlying block device responds with a large delay. What makes me
thinking so is that I use a VM on some cloud provider, and they have
substantial block device latency resulting in permanently high (~20%)
iowait. It spikes occasionally when their cluster is overloaded, and
when that happens, the I/O in my VM may stop and never recover. This is
a rare occasion, but it really happens.
What's worse, so far I've seen such a behaviour with BFQ only. I'm still
testing other schedulers though.
Important note: I have no strict evidences that this is *the* case, thus
I'm asking for some suggestions. My idea is to fire up a local VM and
inject delays to a block device while performing some I/O from within
the VM.
So the question is: how can those delays be injected? Using dm-delay?
Can those delays be random?
Thanks in advance.
--
Oleksandr Natalenko (post-factum)
> Il giorno 21 nov 2019, alle ore 08:13, Oleksandr Natalenko <[email protected]> ha scritto:
>
> Hi Paolo et al.
>
Hi
> I have a strong suspect that something is going wrong when the underlying block device responds with a large delay. What makes me thinking so is that I use a VM on some cloud provider, and they have substantial block device latency resulting in permanently high (~20%) iowait. It spikes occasionally when their cluster is overloaded, and when that happens, the I/O in my VM may stop and never recover. This is a rare occasion, but it really happens.
>
> What's worse, so far I've seen such a behaviour with BFQ only. I'm still testing other schedulers though.
>
> Important note: I have no strict evidences that this is *the* case, thus I'm asking for some suggestions. My idea is to fire up a local VM and inject delays to a block device while performing some I/O from within the VM.
>
> So the question is: how can those delays be injected? Using dm-delay? Can those delays be random?
>
So far I have used scsi_debug [1] for this kind of tests. In my S
suite [2], it boils down to setting SCSI_DEBUG=yes in the S config
file, and then launching any of the benchmarks. Unfortunately, AFAIK
scsi_debug gives you only constant delays; but you can emulate delay
spikes very easily, by changing the delay parameter manually during
the test.
If this option sounds reasonable to you, then I'm willing to help you
for every step.
Thanks,
Paolo
[1] http://sg.danny.cz/sg/sdebug26.html
[2] https://github.com/Algodev-github/S
> Thanks in advance.
>
> --
> Oleksandr Natalenko (post-factum)
> Il giorno 21 nov 2019, alle ore 09:00, Paolo Valente <[email protected]> ha scritto:
>
>
>
>> Il giorno 21 nov 2019, alle ore 08:13, Oleksandr Natalenko <[email protected]> ha scritto:
>>
>> Hi Paolo et al.
>>
>
> Hi
>
>> I have a strong suspect that something is going wrong when the underlying block device responds with a large delay. What makes me thinking so is that I use a VM on some cloud provider, and they have substantial block device latency resulting in permanently high (~20%) iowait. It spikes occasionally when their cluster is overloaded, and when that happens, the I/O in my VM may stop and never recover. This is a rare occasion, but it really happens.
>>
>> What's worse, so far I've seen such a behaviour with BFQ only. I'm still testing other schedulers though.
>>
>> Important note: I have no strict evidences that this is *the* case, thus I'm asking for some suggestions. My idea is to fire up a local VM and inject delays to a block device while performing some I/O from within the VM.
>>
>> So the question is: how can those delays be injected? Using dm-delay? Can those delays be random?
>>
>
> So far I have used scsi_debug [1] for this kind of tests. In my S
> suite [2], it boils down to setting SCSI_DEBUG=yes in the S config
> file, and then launching any of the benchmarks. Unfortunately, AFAIK
> scsi_debug gives you only constant delays; but you can emulate delay
> spikes very easily, by changing the delay parameter manually during
> the test.
>
> If this option sounds reasonable to you, then I'm willing to help you
> for every step.
>
Hi Oleksandr,
Simone (in CC) and I have worked a little bit on reproducing the I/O
freeze you report. Simone made a small change in SCSI_debug, which
makes the latter serve I/O with a highly varying random delay (100ms -
1s), about twice a second.
Then, to generate some fluctuating and heavy I/O, he ran the
comm_startup_lat.sh script of my S suite with SCSI_debug a few times.
Unfortunately, he didn't succeed in reproducing the problem. If you
want, we can send you a patch with his change for SCSI_debug.
Any news on your side?
Thanks,
Simone
> Thanks,
> Paolo
>
> [1] http://sg.danny.cz/sg/sdebug26.html
> [2] https://github.com/Algodev-github/S
>
>> Thanks in advance.
>>
>> --
>> Oleksandr Natalenko (post-factum)
Hello.
On 06.12.2019 17:17, Paolo Valente wrote:
> Simone (in CC) and I have worked a little bit on reproducing the I/O
> freeze you report. Simone made a small change in SCSI_debug, which
> makes the latter serve I/O with a highly varying random delay (100ms -
> 1s), about twice a second.
>
> Then, to generate some fluctuating and heavy I/O, he ran the
> comm_startup_lat.sh script of my S suite with SCSI_debug a few times.
> Unfortunately, he didn't succeed in reproducing the problem. If you
> want, we can send you a patch with his change for SCSI_debug.
>
> Any news on your side?
I was playing with dm-delay in an isolated VM, but so far got no luck.
I'll try to find another way to trigger this (if the bug is still
present in 5.4) and get back to you in case of success.
For me it is a rare occurrence in production, and since I've upgraded to
5.4 and disabled BFQ I haven't seen any at all. At this point I'm not
even sure what I'm looking at. I'll try to re-enable BFQ soon to stress
my production VMs again.
Thank you.
--
Oleksandr Natalenko (post-factum)
Hi.
On 06.12.2019 17:17, Paolo Valente wrote:
> Simone (in CC) and I have worked a little bit on reproducing the I/O
> freeze you report. Simone made a small change in SCSI_debug, which
> makes the latter serve I/O with a highly varying random delay (100ms -
> 1s), about twice a second.
>
> Then, to generate some fluctuating and heavy I/O, he ran the
> comm_startup_lat.sh script of my S suite with SCSI_debug a few times.
> Unfortunately, he didn't succeed in reproducing the problem. If you
> want, we can send you a patch with his change for SCSI_debug.
>
> Any news on your side?
FWIW, I guess I'm safe to exclude BFQ at the moment since I've
encountered a very similar issue without having BFQ enabled.
Also, I think this might be unrelated to the block layer at all. I
suspect there's some race between MADV_MERGEABLE and MADV_DONTNEED since
this is what's hammering the affected tasks and what I see from the call
traces.
I'll investigate further and probably talk to MM people instead. Sorry
for the noise.
--
Oleksandr Natalenko (post-factum)