client/server test to write to single file from all clients

client/server test to write to single file from all clients

[Gurvinder Singh]

Hi,

First of all thanks for such a nice tool to do IO testing. I am
wondering if it is possible to test accessing a single file from
multiple clients to test distributed file system behavior.

The reason for this test is to see if it has different behavior than
having numjobs > 1 accessing same file running on same host.

Any thoughts on this are welcome.

Thanks,
Gurvinder

Re: client/server test to write to single file from all clients

[Jens Axboe]

On 05/01/2016 01:11 PM, Gurvinder Singh wrote:
> Hi,
>
> First of all thanks for such a nice tool to do IO testing. I am
> wondering if it is possible to test accessing a single file from
> multiple clients to test distributed file system behavior.
>
> The reason for this test is to see if it has different behavior than
> having numjobs > 1 accessing same file running on same host.

Sure, you can do that. Fio doesn't really care about whether the file is 
shared or not. Obviously some things like verification might not work 
properly, but just reading and writing to a shared file, fio doesn't 
care if it's local or not.


-- 
Jens Axboe

Re: client/server test to write to single file from all clients

[Gurvinder Singh]

[-- Attachment #1: Type: text/plain, Size: 1017 bytes --]

Thanks Jens for the reply. If FIO supports writing to same file in
client/server mode then what should be filename/filename_pattern should be.
As in currently specifying filename it appends hostname to file name and
thus each client ends up writing to its own file.

- Gurvinder
On May 3, 2016 3:15 AM, "Jens Axboe" <axboe@kernel.dk> wrote:

> On 05/01/2016 01:11 PM, Gurvinder Singh wrote:
>
>> Hi,
>>
>> First of all thanks for such a nice tool to do IO testing. I am
>> wondering if it is possible to test accessing a single file from
>> multiple clients to test distributed file system behavior.
>>
>> The reason for this test is to see if it has different behavior than
>> having numjobs > 1 accessing same file running on same host.
>>
>
> Sure, you can do that. Fio doesn't really care about whether the file is
> shared or not. Obviously some things like verification might not work
> properly, but just reading and writing to a shared file, fio doesn't care
> if it's local or not.
>
>
> --
> Jens Axboe
>
>

[-- Attachment #2: Type: text/html, Size: 1442 bytes --]

Re: client/server test to write to single file from all clients

[Gurvinder Singh]

Reposting it as gmail android client inserted html and it got rejected
from mailing list.
-----------------------------
Thanks Jens for the reply. As FIO supports writing to same file in
client/server mode then what should be filename/filename_pattern should be.
As currently specifying filename it appends hostname/IP to file name and
thus each client ends up writing to its own file.

-------------------------------
- Gurvinder

On 05/03/2016 03:15 AM, Jens Axboe wrote:
> On 05/01/2016 01:11 PM, Gurvinder Singh wrote:
>> Hi,
>>
>> First of all thanks for such a nice tool to do IO testing. I am
>> wondering if it is possible to test accessing a single file from
>> multiple clients to test distributed file system behavior.
>>
>> The reason for this test is to see if it has different behavior than
>> having numjobs > 1 accessing same file running on same host.
> 
> Sure, you can do that. Fio doesn't really care about whether the file is
> shared or not. Obviously some things like verification might not work
> properly, but just reading and writing to a shared file, fio doesn't
> care if it's local or not.
> 
> 

Re: client/server test to write to single file from all clients

[Sitsofe Wheeler]

On 3 May 2016 at 06:34, Gurvinder Singh <gurvindersinghdahiya@gmail.com> wrote:
> Reposting it as gmail android client inserted html and it got rejected
> from mailing list.
> -----------------------------
> Thanks Jens for the reply. As FIO supports writing to same file in
> client/server mode then what should be filename/filename_pattern should be.
> As currently specifying filename it appends hostname/IP to file name and
> thus each client ends up writing to its own file.

Can you post your jobfile here?

-- 
Sitsofe | http://sucs.org/~sits/

Re: client/server test to write to single file from all clients

[Ben England]

Gurvinder, 

you are correct, there is a problem at present getting fio threads on multiple clients to share files using the --client option, because it does insert the IP addr of the client in the pathname.  I think the current behavior is a reasonable default behavior - we want "embarrassingly parallel" behavior normally.  For example, fio threads within a single client run on separate files by default.  But it would be useful to be able to override this and have clients share files - this option would be useful for testing distributed filesystems for example.   Perhaps a "--shared-across-clients" option, valid only when --client specified, would be useful in this case?  This would just suppress the insertion of IP address into the filename.  It should be feasible to do this, if people agree that this is what they want. 

Jens' comment about being unable to verify or predict file contents does apply, if writes are done this way.  So you wouldn't really have a way to know if your writes or reads succeeded.   To make this possible, I suspect fio would need to serialize writes from different clients to the same byte range within a file, I don't know if this can be easily done.

-ben

----- Original Message -----
> From: "Sitsofe Wheeler" <sitsofe@gmail.com>
> To: "Gurvinder Singh" <gurvindersinghdahiya@gmail.com>
> Cc: "Jens Axboe" <axboe@kernel.dk>, fio@vger.kernel.org
> Sent: Tuesday, May 3, 2016 2:18:04 PM
> Subject: Re: client/server test to write to single file from all clients
> 
> On 3 May 2016 at 06:34, Gurvinder Singh <gurvindersinghdahiya@gmail.com>
> wrote:
> > Reposting it as gmail android client inserted html and it got rejected
> > from mailing list.
> > -----------------------------
> > Thanks Jens for the reply. As FIO supports writing to same file in
> > client/server mode then what should be filename/filename_pattern should be.
> > As currently specifying filename it appends hostname/IP to file name and
> > thus each client ends up writing to its own file.
> 
> Can you post your jobfile here?
> 
> --
> Sitsofe | http://sucs.org/~sits/
> --
> To unsubscribe from this list: send the line "unsubscribe fio" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 

Re: client/server test to write to single file from all clients

[Jens Axboe]

On 05/22/2016 04:25 PM, Ben England wrote:
> Gurvinder,
>
> you are correct, there is a problem at present getting fio threads on
> multiple clients to share files using the --client option, because it
> does insert the IP addr of the client in the pathname.  I think the
> current behavior is a reasonable default behavior - we want
> "embarrassingly parallel" behavior normally.  For example, fio threads
> within a single client run on separate files by default.  But it would
> be useful to be able to override this and have clients share files -
> this option would be useful for testing distributed filesystems for
> example.   Perhaps a "--shared-across-clients" option, valid only when
> --client specified, would be useful in this case?  This would just
> suppress the insertion of IP address into the filename.  It should be
> feasible to do this, if people agree that this is what they want.

Just add an option that default to on to retain the existing behavior,
but allows you to turn off the IP prefixing. I'm open to suggestions. It
only happens with a directory prefix, which is indeed a bit odd.o

http://git.kernel.dk/cgit/fio/commit/?id=922a5be82ca50f7c89a1f9c4b84757ba8636e2d1

-- 
Jens Axboe

Re: client/server test to write to single file from all clients

[Gurvinder Singh]

On 05/23/2016 12:25 AM, Ben England wrote:
> Gurvinder, 
> 
> you are correct, there is a problem at present getting fio threads on multiple clients to share files using the --client option, because it does insert the IP addr of the client in the pathname.  I think the current behavior is a reasonable default behavior - we want "embarrassingly parallel" behavior normally.  For example, fio threads within a single client run on separate files by default.  But it would be useful to be able to override this and have clients share files - this option would be useful for testing distributed filesystems for example.   Perhaps a "--shared-across-clients" option, valid only when --client specified, would be useful in this case?  This would just suppress the insertion of IP address into the filename.  It should be feasible to do this, if people agree that this is what they want. 
> 
Thanks Ben for confirming the issue. Yeah the original idea was to test
the distributed file system with multiple clients and writing to same
file. I do understand the verification part is hard to implement in this
case. But adding an option which Jens just did is a good start.

Thanks Jens for fixing it.

- Gurvinder
> Jens' comment about being unable to verify or predict file contents does apply, if writes are done this way.  So you wouldn't really have a way to know if your writes or reads succeeded.   To make this possible, I suspect fio would need to serialize writes from different clients to the same byte range within a file, I don't know if this can be easily done.
> 
> -ben
> 
> ----- Original Message -----
>> From: "Sitsofe Wheeler" <sitsofe@gmail.com>
>> To: "Gurvinder Singh" <gurvindersinghdahiya@gmail.com>
>> Cc: "Jens Axboe" <axboe@kernel.dk>, fio@vger.kernel.org
>> Sent: Tuesday, May 3, 2016 2:18:04 PM
>> Subject: Re: client/server test to write to single file from all clients
>>
>> On 3 May 2016 at 06:34, Gurvinder Singh <gurvindersinghdahiya@gmail.com>
>> wrote:
>>> Reposting it as gmail android client inserted html and it got rejected
>>> from mailing list.
>>> -----------------------------
>>> Thanks Jens for the reply. As FIO supports writing to same file in
>>> client/server mode then what should be filename/filename_pattern should be.
>>> As currently specifying filename it appends hostname/IP to file name and
>>> thus each client ends up writing to its own file.
>>
>> Can you post your jobfile here?
>>
>> --
>> Sitsofe | http://sucs.org/~sits/
>> --
>> To unsubscribe from this list: send the line "unsubscribe fio" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>