Re: Wheezy regressions on QNAP TS-419P+: port issue?
On Thu, Sep 29, 2011 at 2:42 PM, Arnaud Patard
<arnaud.patard@rtp-net.org> wrote:
> Jari Kirma <jari@kirma.fi> writes:
>
> Hi,
>
>> I recently installed Debian Squeeze on a new TS-419P+, and it ran
>> nicely. So, to make it more interesting, I upgraded to Wheeze, and now
>> I'm seeing a performance/responsiveness regression whose cause I can't
>> quite figure out.
>>
>> This has following symptoms:
>>
>> - Big-file dd with large block size (over 16 kiB or so, but especially
>> like 1 MiB), from both RAID-5 backed files and bare disks performs
>> much slower (30-60%!) than before.
>> - dd with small blocks (4 kiB would seem to be close to sweet spot)
>> performs roughly as expected, but slightly slower when served from
>> RAID-5 (maybe 25% difference at most).
>> - SSH and even serial console interaction on other shells feels much
>> choppier than it used to be according to my memory with Squeeze while
>> I was doing these benchmarks.
>> - System is fine without this disk activity (I didn't try if network
>> activity would have a similar effect, but just busylooping processes
>> cause no trouble).
>>
>> Do you have a chance to try this on Kirkwood system, preferably with
>> Marvell 6282, not 6281? Somehow I'm feeling that
>
> iirc, 6282 and 6281 are very similar beasts so it should not matter a
> lot if one is using 6281 or 6282
>
>> linux-image-3.0.0-1-kirkwood kernel is missing interrupts, or
>> something. There is no indication of this on dmesg, though.
>
> There has not been a lot of changes on kirkwood specific parts so your
> issue may not be kirkwood specific. If it's possible, would be
> interesting to test 3.1-rc7 from experimental on your ts419p+. It's just
> finished to build so it should be available soon.
3.1-rc7 indeed helped! I wonder what was the cause of this (most
likely non-arm-related) regression, but on 3.0, performance of block
reads was reasonable for block sizes of 1-96 kiB, started to
noticeably drop at 128 kiB and became downright abysmal (less than 40%
of the peak performance) around 256 kiB. With 3.1-rc7, slight
degradation (as earlier with squeeze) can be noticed past 128 kiB, but
transfer rates still stay around 85-95% bracket. Everything seems back
to normal now. :)
Reply to: