[lustre-discuss] Asynchronous POSIX I/O with Lustre

[Dilger, Andreas]

On 2016/05/12, 09:11, "Sebastian Rettenberger" <rettenbs at in.tum.de> wrote:

>Hi,
>
>thanks for the feedback.
>
>I found out a few things (especially for Linux) that might be interesting:
>
>- glibc implements aio_read/aio_write in user-space using threads. Thus,
>they are always asynchronous but come with an overhead.
>
>- The Linux kernel provides non-portable native asynchronous I/O through
>libaio (io_submit).
>
>I tested native AIO function of Linux on Ext4, NFS, Lustre and GPFS.
>Only NFS performed the operation asynchronously and only with O_DIRECT.
>On all other file systems, the operation was turned in a synchronous
>operation.

Could you please file a ticket in Jira and share what test tool you are
using to check the aio functionality?  That would at least allow someone
to investigate ow this is implemented and where it is getting blocked.
I can't promise someone will work on it any time soon, but at least we
would have somewhere to start, since I don't think this is a common IO
interface except with database workloads.

Cheers, Andreas


>On 05/12/2016 12:23 AM, Dilger, Andreas wrote:
>> On 2016/05/10, 04:30, "Sebastian Rettenberger" <rettenbs at in.tum.de>
>>wrote:
>>
>>> Hi,
>>>
>>> I am currently investigating in asynchronous I/O capability of
>>>different
>>> file systems. I wounder whether Lustre supports aio_write and aio_read
>>>from the POSIX standard.
>>>
>>> I found this bug report:
>>> https://bugzilla.lustre.org/show_bug.cgi?id=11719
>>> and this archived mail:
>>> 
>>>http://lists.lustre.org/htdig.cgi/lustre-discuss-lustre.org/2012-May/011
>>>08
>>> 7.html
>>>
>>> Does anybody has any resent information on this? Is this docuemented in
>>> the Lustre manual?
>>
>> The Lustre client does use the kernel ->aio_read() and ->aio_write()
>> operations, but I don't know whether anyone has actually tested whether
>>the
>> IO operations will commit asynchronously or not.  If you have a test
>>tool
>> do verify this, it would definitely be interesting to know the results.
>>
>> Cheers, Andreas
>>
>
>-- 
>Sebastian Rettenberger, M.Sc.
>Technische Universität München
>Department of Informatics
>Chair of Scientific Computing
>Boltzmannstrasse 3, 85748 Garching, Germany
>http://www5.in.tum.de/
>
>


Cheers, Andreas
-- 
Andreas Dilger

Lustre Principal Architect
Intel High Performance Data Division