Re: [Qemu-devel] [RFC 0/3] aio: experimental virtio-blk polling mode

[Karl Rister]

On 11/15/2016 04:32 AM, Stefan Hajnoczi wrote:
> On Mon, Nov 14, 2016 at 09:52:00PM +0100, Paolo Bonzini wrote:
>> On 14/11/2016 21:12, Karl Rister wrote:
>>>                  256    46,929
>>>                  512    35,627
>>>                1,024    46,477
>>>                2,000    35,247
>>>                2,048    46,322
>>>                4,000    46,540
>>>                4,096    46,368
>>>                8,000    47,054
>>>                8,192    46,671
>>>               16,000    46,466
>>>               16,384    32,504
>>>               32,000    20,620
>>>               32,768    20,807
>>
>> Huh, it breaks down exactly when it should start going faster
>> (10^9/46000 = ~21000).
> 
> Could it be because we're not breaking the polling loop for BHs, new
> timers, or aio_notify()?
> 
> Once that is fixed polling should achieve maximum performance when
> QEMU_AIO_MAX_POLL_NS is at least as long as the duration of a request.
> 
> This is logical if there are enough pinned CPUs so the polling thread
> can run flat out.
> 

I removed all the pinning and restored the guest to a "normal"
configuration.

QEMU_AIO_POLL_MAX_NS      IOPs
               unset    25,553
                   1    28,684
                   2    38,213
                   4    29,413
                   8    38,612
                  16    30,578
                  32    30,145
                  64    41,637
                 128    28,554
                 256    29,661
                 512    39,178
               1,024    29,644
               2,048    37,190
               4,096    29,838
               8,192    38,581
              16,384    37,793
              32,768    20,332
              65,536    35,755

-- 
Karl Rister <address@hidden>