[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH] linux-aio: keep processing events if MAX_EVENTS
From: |
Paolo Bonzini |
Subject: |
Re: [Qemu-devel] [PATCH] linux-aio: keep processing events if MAX_EVENTS reached |
Date: |
Tue, 12 Jul 2016 18:58:22 +0200 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.1.1 |
On 12/07/2016 18:34, Roman Penyaev wrote:
> On Tue, Jul 12, 2016 at 4:59 PM, Paolo Bonzini <address@hidden> wrote:
>> On 12/07/2016 16:12, Roman Penyaev wrote:
>>> But what is the most important thing here is, that reverting
>>> "linux-aio: Cancel BH if not needed" brings these numbers:
>>>
>>> READ: io=56362MB, aggrb=1878.4MB/s, minb=1878.4MB/s,
>>> maxb=1878.4MB/s, mint=30007msec, maxt=30007msec
>>> WRITE: io=56255MB, aggrb=1874.8MB/s, minb=1874.8MB/s,
>>> maxb=1874.8MB/s, mint=30007msec, maxt=30007msec
>>>
>>> So, it seems to me that "linux-aio: Cancel BH if not needed" introduces
>>> regression.
>>
>> It's possible that this depends on the workload. But, thanks for
>> measuring it. It's not a small effect (~5%). We should consider indeed
>> reverting the patch.
>
> Without any doubts it depends. But I can't say that my fio config
> is something special
Yup, I wasn't expecting anything special.
>>> - unsigned int n;
>>> + unsigned int inqueue;
>>> + unsigned int inflight;
>>
>> Please use in_queue (or queue_length) and in_flight.
>
> Yeah, but that was just a quick exercise.
> If Stefan or you want this as a normal patch, I will resend
> with all tweaks made.
Yes, please do send it.
Paolo