Re: [Qemu-devel] [PATCH v2 1/5] linux-aio: queue requests that cannot be submitted

[Paolo Bonzini <pbonzini@redhat.com>]

On 17/12/2014 13:34, Paolo Bonzini wrote:
> 
> 
> On 16/12/2014 21:26, Paolo Bonzini wrote:
>>
>>> I could reproduce this very well on a random OS image that I had around.
>>>  This is raw over XFS over dm-crypt, and the image is about 75% sparse
>>> (8.2G used over 35G).  I only get 1-2%, but still it's visible.
>>>
>>> However I can hardly reproduce it when using a partition directly:
>>>
>>>          old    new
>>> mean     9.9565 9.9636      (+0.07%)
>>> stddev   0.0405 0.0537
>>> min      9.871  9.867
>>> median   9.973  9.971
>>> max      10.01  10.053
>>> count    20     20
>>>
>>> I haven't tried removing layers (e.g. fully-allocated XFS image without
>>> dm-crypt).
>>
>> Could not reproduce it with a fully-allocated XFS image, on the contrary
>> the patched QEMU is a bit faster:
>>
>>         old      new
>> mean    14.83325 14.82660
>> stddev  0.016930 0.010328
>> min     14.819   14.818
>> max     14.854   14.883
>> median  14.8225  14.8255
>> count   20       20
> 
> Same for 50% sparse XFS image, patched QEMU a bit faster:
> 
>         old     new
> Mean	8.31285 8.30625
> stddev	0.03690 0.03333
> min     8.277   8.276
> max     8.378   8.382
> median	8.292   8.2905
> count   20      20

Today I cannot reproduce it even on the original testcase:

       old     new
mean   9.64175 9.60935
stddev 0.15211 0.11117
min    9.445   9.454
max    10.102  9.835
median 9.64    9.6205
count  20      20

Some notes:

1) do you have Fam's io_get_events patch?  I don't, and the tests
probably should be done with it.

2) "qemu-img bench" probably could also be changed to use AioContext
directly instead of going through GMainLoop.  That said, there is some
low-hanging fruit for GMainLoop that I'll shortly send a patch for.

3) bench_cb does this:

        b->sector += b->bufsize;

Should it shift right by >> BDRV_SECTOR_BITS?  Or is it intended to do
the equivalent of a random read benchmark?

Paolo